Daily Papers

Though safety alignment has been applied to most large language models (LLMs), LLM service providers generally deploy a subsequent moderation as the external safety guardrail in real-world products. Existing moderators mainly practice a conventional full detection, which determines the harmfulness based on the complete LLM output, causing high service latency. Recent works pay more attention to partial detection where moderators oversee the generation midway and early stop the output if harmfulness is detected, but they directly apply moderators trained with the full detection paradigm to incomplete outputs, introducing a training-inference gap that lowers the performance. In this paper, we explore how to form a data-and-model solution that natively supports partial detection. For the data, we construct FineHarm, a dataset consisting of 29K prompt-response pairs with fine-grained annotations to provide reasonable supervision for token-level training. Then, we propose the streaming content monitor, which is trained with dual supervision of response- and token-level labels and can follow the output stream of LLM to make a timely judgment of harmfulness. Experiments show that SCM gains 0.95+ in macro F1 score that is comparable to full detection, by only seeing the first 18% of tokens in responses on average. Moreover, the SCM can serve as a pseudo-harmfulness annotator for improving safety alignment and lead to a higher harmlessness score than DPO.

Large Reasoning Models (LRMs) have become powerful tools for complex problem solving, but their structured reasoning pathways can lead to unsafe outputs when exposed to harmful prompts. Existing safety alignment methods reduce harmful outputs but can degrade reasoning depth, leading to significant trade-offs in complex, multi-step tasks, and remain vulnerable to sophisticated jailbreak attacks. To address this, we introduce SAFEPATH, a lightweight alignment method that fine-tunes LRMs to emit a short, 8-token Safety Primer at the start of their reasoning, in response to harmful prompts, while leaving the rest of the reasoning process unsupervised. Empirical results across multiple benchmarks indicate that SAFEPATH effectively reduces harmful outputs while maintaining reasoning performance. Specifically, SAFEPATH reduces harmful responses by up to 90.0% and blocks 83.3% of jailbreak attempts in the DeepSeek-R1-Distill-Llama-8B model, while requiring 295.9x less compute than Direct Refusal and 314.1x less than SafeChain. We further introduce a zero-shot variant that requires no fine-tuning. In addition, we provide a comprehensive analysis of how existing methods in LLMs generalize, or fail, when applied to reasoning-centric models, revealing critical gaps and new directions for safer AI.

Jailbreak attacks bypass the guardrails of large language models to produce harmful outputs. In this paper, we ask whether the model outputs produced by existing jailbreaks are actually useful. For example, when jailbreaking a model to give instructions for building a bomb, does the jailbreak yield good instructions? Since the utility of most unsafe answers (e.g., bomb instructions) is hard to evaluate rigorously, we build new jailbreak evaluation sets with known ground truth answers, by aligning models to refuse questions related to benign and easy-to-evaluate topics (e.g., biology or math). Our evaluation of eight representative jailbreaks across five utility benchmarks reveals a consistent drop in model utility in jailbroken responses, which we term the jailbreak tax. For example, while all jailbreaks we tested bypass guardrails in models aligned to refuse to answer math, this comes at the expense of a drop of up to 92% in accuracy. Overall, our work proposes the jailbreak tax as a new important metric in AI safety, and introduces benchmarks to evaluate existing and future jailbreaks. We make the benchmark available at https://github.com/ethz-spylab/jailbreak-tax

Evaluating jailbreak attacks is challenging when prompts are not overtly harmful or fail to induce harmful outputs. Unfortunately, many existing red-teaming datasets contain such unsuitable prompts. To evaluate attacks accurately, these datasets need to be assessed and cleaned for maliciousness. However, existing malicious content detection methods rely on either manual annotation, which is labor-intensive, or large language models (LLMs), which have inconsistent accuracy in harmful types. To balance accuracy and efficiency, we propose a hybrid evaluation framework named MDH (Malicious content Detection based on LLMs with Human assistance) that combines LLM-based annotation with minimal human oversight, and apply it to dataset cleaning and detection of jailbroken responses. Furthermore, we find that well-crafted developer messages can significantly boost jailbreak success, leading us to propose two new strategies: D-Attack, which leverages context simulation, and DH-CoT, which incorporates hijacked chains of thought. The Codes, datasets, judgements, and detection results will be released in github repository: https://github.com/AlienZhang1996/DH-CoT.

Large language models (LLMs) exhibit exceptional capabilities across various tasks but also pose risks by generating harmful content. Existing safety mechanisms, while improving model safety, often lead to overly cautious behavior and fail to fully leverage LLMs' internal cognitive processes. Inspired by humans' reflective thinking capability, we first show that LLMs can similarly perform internal assessments about safety in their internal states. Building on this insight, we propose SafeSwitch, a dynamic framework that regulates unsafe outputs by utilizing the prober-based internal state monitor that actively detects harmful intentions, and activates a safety head that leads to safer and more conservative responses only when necessary. SafeSwitch reduces harmful outputs by approximately 80% on harmful queries while maintaining strong utility, reaching a Pareto optimal among several methods. Our method is also advantageous over traditional methods in offering more informative, context-aware refusals, and achieves these benefits while only tuning less than 6% of the original parameters. SafeSwitch demonstrates large language models' capacity for self-awareness and reflection regarding safety, offering a promising approach to more nuanced and effective safety controls. Codes for this work are available at https://github.com/Hanpx20/SafeSwitch.

As large language models (LLMs) have increased in their capabilities, so does their potential for dual use. To reduce harmful outputs, produces and vendors of LLMs have used reinforcement learning with human feedback (RLHF). In tandem, LLM vendors have been increasingly enabling fine-tuning of their most powerful models. However, concurrent work has shown that fine-tuning can remove RLHF protections. We may expect that the most powerful models currently available (GPT-4) are less susceptible to fine-tuning attacks. In this work, we show the contrary: fine-tuning allows attackers to remove RLHF protections with as few as 340 examples and a 95% success rate. These training examples can be automatically generated with weaker models. We further show that removing RLHF protections does not decrease usefulness on non-censored outputs, providing evidence that our fine-tuning strategy does not decrease usefulness despite using weaker models to generate training data. Our results show the need for further research on protections on LLMs.

Rapid deployment of vision-language models (VLMs) magnifies safety risks, yet most evaluations rely on artificial images. This study asks: How safe are current VLMs when confronted with meme images that ordinary users share? To investigate this question, we introduce MemeSafetyBench, a 50,430-instance benchmark pairing real meme images with both harmful and benign instructions. Using a comprehensive safety taxonomy and LLM-based instruction generation, we assess multiple VLMs across single and multi-turn interactions. We investigate how real-world memes influence harmful outputs, the mitigating effects of conversational context, and the relationship between model scale and safety metrics. Our findings demonstrate that VLMs show greater vulnerability to meme-based harmful prompts than to synthetic or typographic images. Memes significantly increase harmful responses and decrease refusals compared to text-only inputs. Though multi-turn interactions provide partial mitigation, elevated vulnerability persists. These results highlight the need for ecologically valid evaluations and stronger safety mechanisms.

Deploying Large language models (LLMs) can pose hazards from harmful outputs such as toxic or dishonest speech. Prior work has introduced tools that elicit harmful outputs in order to identify and mitigate these risks. While this is a valuable step toward securing language models, these approaches typically rely on a pre-existing classifier for undesired outputs. This limits their application to situations where the type of harmful behavior is known with precision beforehand. However, this skips a central challenge of red teaming: developing a contextual understanding of the behaviors that a model can exhibit. Furthermore, when such a classifier already exists, red teaming has limited marginal value because the classifier could simply be used to filter training data or model outputs. In this work, we consider red teaming under the assumption that the adversary is working from a high-level, abstract specification of undesired behavior. The red team is expected to refine/extend this specification and identify methods to elicit this behavior from the model. Our red teaming framework consists of three steps: 1) Exploring the model's behavior in the desired context; 2) Establishing a measurement of undesired behavior (e.g., a classifier trained to reflect human evaluations); and 3) Exploiting the model's flaws using this measure and an established red teaming methodology. We apply this approach to red team GPT-2 and GPT-3 models to systematically discover classes of prompts that elicit toxic and dishonest statements. In doing so, we also construct and release the CommonClaim dataset of 20,000 statements that have been labeled by human subjects as common-knowledge-true, common-knowledge-false, or neither. Code is available at https://github.com/thestephencasper/explore_establish_exploit_llms. CommonClaim is available at https://github.com/thestephencasper/common_claim.

The increasing deployment of Large Vision-Language Models (LVLMs) raises safety concerns under potential malicious inputs. However, existing multimodal safety evaluations primarily focus on model vulnerabilities exposed by static image inputs, ignoring the temporal dynamics of video that may induce distinct safety risks. To bridge this gap, we introduce Video-SafetyBench, the first comprehensive benchmark designed to evaluate the safety of LVLMs under video-text attacks. It comprises 2,264 video-text pairs spanning 48 fine-grained unsafe categories, each pairing a synthesized video with either a harmful query, which contains explicit malice, or a benign query, which appears harmless but triggers harmful behavior when interpreted alongside the video. To generate semantically accurate videos for safety evaluation, we design a controllable pipeline that decomposes video semantics into subject images (what is shown) and motion text (how it moves), which jointly guide the synthesis of query-relevant videos. To effectively evaluate uncertain or borderline harmful outputs, we propose RJScore, a novel LLM-based metric that incorporates the confidence of judge models and human-aligned decision threshold calibration. Extensive experiments show that benign-query video composition achieves average attack success rates of 67.2%, revealing consistent vulnerabilities to video-induced attacks. We believe Video-SafetyBench will catalyze future research into video-based safety evaluation and defense strategies.

AI systems can take harmful actions and are highly vulnerable to adversarial attacks. We present an approach, inspired by recent advances in representation engineering, that "short-circuits" models as they respond with harmful outputs. Existing techniques aimed at improving alignment, such as refusal training, are often bypassed. Techniques such as adversarial training try to plug these holes by countering specific attacks. As an alternative to refusal training and adversarial training, short-circuiting directly controls the representations that are responsible for harmful outputs in the first place. Our technique can be applied to both text-only and multimodal language models to prevent the generation of harmful outputs without sacrificing utility -- even in the presence of powerful unseen attacks. Notably, while adversarial robustness in standalone image recognition remains an open challenge, short-circuiting allows the larger multimodal system to reliably withstand image "hijacks" that aim to produce harmful content. Finally, we extend our approach to AI agents, demonstrating considerable reductions in the rate of harmful actions when they are under attack. Our approach represents a significant step forward in the development of reliable safeguards to harmful behavior and adversarial attacks.

Ensuring the safety and harmlessness of Large Language Models (LLMs) has become equally critical as their performance in applications. However, existing safety alignment methods typically suffer from safety-performance trade-offs and the susceptibility to jailbreak attacks, primarily due to their reliance on direct refusals for malicious queries. In this paper, we propose STAIR, a novel framework that integrates SafeTy Alignment with Itrospective Reasoning. We enable LLMs to identify safety risks through step-by-step analysis by self-improving chain-of-thought (CoT) reasoning with safety awareness. STAIR first equips the model with a structured reasoning capability and then advances safety alignment via iterative preference optimization on step-level reasoning data generated using our newly proposed Safety-Informed Monte Carlo Tree Search (SI-MCTS). We further train a process reward model on this data to guide test-time searches for improved responses. Extensive experiments show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies. With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks. Relevant resources in this work are available at https://github.com/thu-ml/STAIR.

Open-source large language models are becoming increasingly available and popular among researchers and practitioners. While significant progress has been made on open-weight models, open training data is a practice yet to be adopted by the leading open-weight models creators. At the same time, there researchers are working to make language models safer. We propose a data curation pipeline to reduce harmful outputs by models trained on public domain data. There are unique challenges to working with public domain data, as these sources differ from web text in both form and content. Many sources are historical documents and are the result of Optical Character Recognition (OCR). Consequently, current state-of-the-art approaches to toxicity filtering are often infeasible or inappropriate for open data models. In this paper, we introduce a new fully open-source pipeline for open-data toxicity filtering. Our contributions are threefold. We create a custom training dataset, ToxicCommons, which is composed of texts which have been classified across five different dimensions (racial/origin-based, gender/sex-based, religious, ability-based discrimination, and violence). We use this dataset to train a custom classifier, Celadon, that can be used to detect toxic content in open data more efficiently at a larger scale. Finally, we describe the balanced approach to content filtration that optimizes safety filtering with respect to the filtered data available for training.

Refusal training on Large Language Models (LLMs) prevents harmful outputs, yet this defense remains vulnerable to both automated and human-crafted jailbreaks. We present a novel LLM-as-red-teamer approach in which a human jailbreaks a refusal-trained LLM to make it willing to jailbreak itself or other LLMs. We refer to the jailbroken LLMs as J_2 attackers, which can systematically evaluate target models using various red teaming strategies and improve its performance via in-context learning from the previous failures. Our experiments demonstrate that Sonnet 3.5 and Gemini 1.5 pro outperform other LLMs as J_2, achieving 93.0% and 91.0% attack success rates (ASRs) respectively against GPT-4o (and similar results across other capable LLMs) on Harmbench. Our work not only introduces a scalable approach to strategic red teaming, drawing inspiration from human red teamers, but also highlights jailbreaking-to-jailbreak as an overlooked failure mode of the safeguard. Specifically, an LLM can bypass its own safeguards by employing a jailbroken version of itself that is willing to assist in further jailbreaking. To prevent any direct misuse with J_2, while advancing research in AI safety, we publicly share our methodology while keeping specific prompting details private.

Text-to-Image (T2I) models have achieved remarkable success in generating visual content from text inputs. Although multiple safety alignment strategies have been proposed to prevent harmful outputs, they often lead to overly cautious behavior -- rejecting even benign prompts -- a phenomenon known as over-refusal that reduces the practical utility of T2I models. Despite over-refusal having been observed in practice, there is no large-scale benchmark that systematically evaluates this phenomenon for T2I models. In this paper, we present an automatic workflow to construct synthetic evaluation data, resulting in OVERT (OVEr-Refusal evaluation on Text-to-image models), the first large-scale benchmark for assessing over-refusal behaviors in T2I models. OVERT includes 4,600 seemingly harmful but benign prompts across nine safety-related categories, along with 1,785 genuinely harmful prompts (OVERT-unsafe) to evaluate the safety-utility trade-off. Using OVERT, we evaluate several leading T2I models and find that over-refusal is a widespread issue across various categories (Figure 1), underscoring the need for further research to enhance the safety alignment of T2I models without compromising their functionality. As a preliminary attempt to reduce over-refusal, we explore prompt rewriting; however, we find it often compromises faithfulness to the meaning of the original prompts. Finally, we demonstrate the flexibility of our generation framework in accommodating diverse safety requirements by generating customized evaluation data adapting to user-defined policies.

Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories -- Knowledge Editing, Model Compression, and Machine Unlearning. Our results from 310 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions. All of our code is public: https://github.com/hartvigsen-group/composable-interventions.

Recent work has shown it is possible to construct adversarial examples that cause an aligned language model to emit harmful strings or perform harmful behavior. Existing attacks work either in the white-box setting (with full access to the model weights), or through transferability: the phenomenon that adversarial examples crafted on one model often remain effective on other models. We improve on prior work with a query-based attack that leverages API access to a remote language model to construct adversarial examples that cause the model to emit harmful strings with (much) higher probability than with transfer-only attacks. We validate our attack on GPT-3.5 and OpenAI's safety classifier; we can cause GPT-3.5 to emit harmful strings that current transfer attacks fail at, and we can evade the safety classifier with nearly 100% probability.

Large Language Models (LLMs) have shown remarkable success in various tasks, but concerns about their safety and the potential for generating malicious content have emerged. In this paper, we explore the power of In-Context Learning (ICL) in manipulating the alignment ability of LLMs. We find that by providing just few in-context demonstrations without fine-tuning, LLMs can be manipulated to increase or decrease the probability of jailbreaking, i.e. answering malicious prompts. Based on these observations, we propose In-Context Attack (ICA) and In-Context Defense (ICD) methods for jailbreaking and guarding aligned language model purposes. ICA crafts malicious contexts to guide models in generating harmful outputs, while ICD enhances model robustness by demonstrations of rejecting to answer harmful prompts. Our experiments show the effectiveness of ICA and ICD in increasing or reducing the success rate of adversarial jailbreaking attacks. Overall, we shed light on the potential of ICL to influence LLM behavior and provide a new perspective for enhancing the safety and alignment of LLMs.

The safety and alignment of Large Language Models (LLMs) are critical for their responsible deployment. Current evaluation methods predominantly focus on identifying and preventing overtly harmful outputs. However, they often fail to address a more insidious failure mode: models that produce benign-appearing outputs while operating on malicious or deceptive internal reasoning. This vulnerability, often triggered by sophisticated system prompt injections, allows models to bypass conventional safety filters, posing a significant, underexplored risk. To address this gap, we introduce the Deceptive Reasoning Exposure Suite (D-REX), a novel dataset designed to evaluate the discrepancy between a model's internal reasoning process and its final output. D-REX was constructed through a competitive red-teaming exercise where participants crafted adversarial system prompts to induce such deceptive behaviors. Each sample in D-REX contains the adversarial system prompt, an end-user's test query, the model's seemingly innocuous response, and, crucially, the model's internal chain-of-thought, which reveals the underlying malicious intent. Our benchmark facilitates a new, essential evaluation task: the detection of deceptive alignment. We demonstrate that D-REX presents a significant challenge for existing models and safety mechanisms, highlighting the urgent need for new techniques that scrutinize the internal processes of LLMs, not just their final outputs.

As vision-language models (VLMs) gain prominence, their multimodal interfaces also introduce new safety vulnerabilities, making the safety evaluation challenging and critical. Existing red-teaming efforts are either restricted to a narrow set of adversarial patterns or depend heavily on manual engineering, lacking scalable exploration of emerging real-world VLM vulnerabilities. To bridge this gap, we propose ARMs, an adaptive red-teaming agent that systematically conducts comprehensive risk assessments for VLMs. Given a target harmful behavior or risk definition, ARMs automatically optimizes diverse red-teaming strategies with reasoning-enhanced multi-step orchestration, to effectively elicit harmful outputs from target VLMs. We propose 11 novel multimodal attack strategies, covering diverse adversarial patterns of VLMs (e.g., reasoning hijacking, contextual cloaking), and integrate 17 red-teaming algorithms into ARMs via model context protocol (MCP). To balance the diversity and effectiveness of the attack, we design a layered memory with an epsilon-greedy attack exploration algorithm. Extensive experiments on instance- and policy-based benchmarks show that ARMs achieves SOTA attack success rates, exceeding baselines by an average of 52.1% and surpassing 90% on Claude-4-Sonnet. We show that the diversity of red-teaming instances generated by ARMs is significantly higher, revealing emerging vulnerabilities in VLMs. Leveraging ARMs, we construct ARMs-Bench, a large-scale multimodal safety dataset comprising over 30K red-teaming instances spanning 51 diverse risk categories, grounded in both real-world multimodal threats and regulatory risks. Safety fine-tuning with ARMs-Bench substantially improves the robustness of VLMs while preserving their general utility, providing actionable guidance to improve multimodal safety alignment against emerging threats.

AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat, a collection of instruction fine-tuned large language models, they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. However, it remains unclear how well safety training guards against model misuse when attackers have access to model weights. We explore the robustness of safety training in language models by subversively fine-tuning the public weights of Llama 2-Chat. We employ low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than $200 per model and using only one GPU, we successfully undo the safety training of Llama 2-Chat models of sizes 7B, 13B, and 70B. Specifically, our fine-tuning technique significantly reduces the rate at which the model refuses to follow harmful instructions. We achieve a refusal rate below 1% for our 70B Llama 2-Chat model on two refusal benchmarks. Our fine-tuning method retains general performance, which we validate by comparing our fine-tuned models against Llama 2-Chat across two benchmarks. Additionally, we present a selection of harmful outputs produced by our models. While there is considerable uncertainty about the scope of risks from current models, it is likely that future models will have significantly more dangerous capabilities, including the ability to hack into critical infrastructure, create dangerous bio-weapons, or autonomously replicate and adapt to new environments. We show that subversive fine-tuning is practical and effective, and hence argue that evaluating risks from fine-tuning should be a core part of risk assessments for releasing model weights.

Safety fine-tuning algorithms are commonly used to fine-tune language models to reduce harmful outputs, but the exact internal mechanisms of how those models achieve this remain unclear. In studying direct preference optimisation (DPO) for toxicity reduction, current explanations claim that DPO works by dampening the most toxic MLP neurons to learn an offset to avert toxic regions in the residual stream. However, by ablating the most toxic neurons and applying activation patching, we find this explanation incomplete. By projecting neuron activation changes onto a toxicity probe, we find that only 31.8\% of toxicity reduction comes from dampened toxic neurons. Instead, DPO reduces toxicity by accumulating effects across multiple neuron groups, both reducing writing in the toxic direction and promoting anti-toxicity in the residual stream. Moreover, DPO gives noisy adjustments to neuron activations, with many neurons actually increasing toxicity. This indicates that DPO is a balancing process between opposing neuron effects to achieve toxicity reduction.

Large language models, despite extensive alignment with human values and ethical principles, remain vulnerable to sophisticated jailbreak attacks that exploit their reasoning abilities. Existing safety measures often detect overt malicious intent but fail to address subtle, reasoning-driven vulnerabilities. In this work, we introduce POATE (Polar Opposite query generation, Adversarial Template construction, and Elaboration), a novel jailbreak technique that harnesses contrastive reasoning to provoke unethical responses. POATE crafts semantically opposing intents and integrates them with adversarial templates, steering models toward harmful outputs with remarkable subtlety. We conduct extensive evaluation across six diverse language model families of varying parameter sizes to demonstrate the robustness of the attack, achieving significantly higher attack success rates (~44%) compared to existing methods. To counter this, we propose Intent-Aware CoT and Reverse Thinking CoT, which decompose queries to detect malicious intent and reason in reverse to evaluate and reject harmful responses. These methods enhance reasoning robustness and strengthen the model's defense against adversarial exploits.

Multimodal Language Models (MMLMs) typically undergo post-training alignment to prevent harmful content generation. However, these alignment stages focus primarily on the assistant role, leaving the user role unaligned, and stick to a fixed input prompt structure of special tokens, leaving the model vulnerable when inputs deviate from these expectations. We introduce Role-Modality Attacks (RMA), a novel class of adversarial attacks that exploit role confusion between the user and assistant and alter the position of the image token to elicit harmful outputs. Unlike existing attacks that modify query content, RMAs manipulate the input structure without altering the query itself. We systematically evaluate these attacks across multiple Vision Language Models (VLMs) on eight distinct settings, showing that they can be composed to create stronger adversarial prompts, as also evidenced by their increased projection in the negative refusal direction in the residual stream, a property observed in prior successful attacks. Finally, for mitigation, we propose an adversarial training approach that makes the model robust against input prompt perturbations. By training the model on a range of harmful and benign prompts all perturbed with different RMA settings, it loses its sensitivity to Role Confusion and Modality Manipulation attacks and is trained to only pay attention to the content of the query in the input prompt structure, effectively reducing Attack Success Rate (ASR) while preserving the model's general utility.

Generating a Chain of Thought (CoT) has been shown to consistently improve large language model (LLM) performance on a wide range of NLP tasks. However, prior work has mainly focused on logical reasoning tasks (e.g. arithmetic, commonsense QA); it remains unclear whether improvements hold for more diverse types of reasoning, especially in socially situated contexts. Concretely, we perform a controlled evaluation of zero-shot CoT across two socially sensitive domains: harmful questions and stereotype benchmarks. We find that zero-shot CoT reasoning in sensitive domains significantly increases a model's likelihood to produce harmful or undesirable output, with trends holding across different prompt formats and model variants. Furthermore, we show that harmful CoTs increase with model size, but decrease with improved instruction following. Our work suggests that zero-shot CoT should be used with caution on socially important tasks, especially when marginalized groups or sensitive topics are involved.

As large language models (LLMs) become more powerful and are deployed more autonomously, it will be increasingly important to prevent them from causing harmful outcomes. Researchers have investigated a variety of safety techniques for this purpose, e.g. using models to review the outputs of other models, or red-teaming techniques to surface subtle failure modes. However, researchers have not evaluated whether such techniques still ensure safety if the model is itself intentionally trying to subvert them. In this paper, we develop and evaluate pipelines of safety techniques ("protocols") that are robust to intentional subversion. We investigate a scenario in which we want to solve a sequence of programming problems, using access to a powerful but untrusted model (in our case, GPT-4), access to a less powerful trusted model (in our case, GPT-3.5), and limited access to high-quality trusted labor. We investigate protocols that aim to never submit solutions containing backdoors, which we operationalize here as logical errors that are not caught by test cases. We investigate a range of protocols and test each against strategies that the untrusted model could use to subvert them. One protocol is what we call trusted editing. This protocol first asks GPT-4 to write code, and then asks GPT-3.5 to rate the suspiciousness of that code. If the code is below some suspiciousness threshold, it is submitted. Otherwise, GPT-3.5 edits the solution to remove parts that seem suspicious and then submits the edited code. Another protocol is untrusted monitoring. This protocol asks GPT-4 to write code, and then asks another instance of GPT-4 whether the code is backdoored, using various techniques to prevent the GPT-4 instances from colluding. These protocols improve substantially on simple baselines.

We describe our early efforts to red team language models in order to simultaneously discover, measure, and attempt to reduce their potentially harmful outputs. We make three main contributions. First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF). We find that the RLHF models are increasingly difficult to red team as they scale, and we find a flat trend with scale for the other model types. Second, we release our dataset of 38,961 red team attacks for others to analyze and learn from. We provide our own analysis of the data and find a variety of harmful outputs, which range from offensive language to more subtly harmful non-violent unethical outputs. Third, we exhaustively describe our instructions, processes, statistical methodologies, and uncertainty about red teaming. We hope that this transparency accelerates our ability to work together as a community in order to develop shared norms, practices, and technical standards for how to red team language models.

Large Language Models (LLMs) are usually aligned with "human values/preferences" to prevent harmful output. Discussions around the alignment of Large Language Models (LLMs) generally focus on preventing harmful outputs. However, in this paper, we argue that in health-related queries, over-alignment-leading to overly cautious responses-can itself be harmful, especially for people with anxiety and obsessive-compulsive disorder (OCD). This is not only unethical but also dangerous to the user, both mentally and physically. We also showed qualitative results that some LLMs exhibit varying degrees of alignment. Finally, we call for the development of LLMs with stronger reasoning capabilities that provide more tailored and nuanced responses to health queries. Warning: This paper contains materials that could trigger health anxiety or OCD.

A high volume of recent ML security literature focuses on attacks against aligned large language models (LLMs). These attacks may extract private information or coerce the model into producing harmful outputs. In real-world deployments, LLMs are often part of a larger agentic pipeline including memory systems, retrieval, web access, and API calling. Such additional components introduce vulnerabilities that make these LLM-powered agents much easier to attack than isolated LLMs, yet relatively little work focuses on the security of LLM agents. In this paper, we analyze security and privacy vulnerabilities that are unique to LLM agents. We first provide a taxonomy of attacks categorized by threat actors, objectives, entry points, attacker observability, attack strategies, and inherent vulnerabilities of agent pipelines. We then conduct a series of illustrative attacks on popular open-source and commercial agents, demonstrating the immediate practical implications of their vulnerabilities. Notably, our attacks are trivial to implement and require no understanding of machine learning.

Current Vision Language Models (VLMs) remain vulnerable to malicious prompts that induce harmful outputs. Existing safety benchmarks for VLMs primarily rely on automated evaluation methods, but these methods struggle to detect implicit harmful content or produce inaccurate evaluations. Therefore, we found that existing benchmarks have low levels of harmfulness, ambiguous data, and limited diversity in image-text pair combinations. To address these issues, we propose the ELITE benchmark, a high-quality safety evaluation benchmark for VLMs, underpinned by our enhanced evaluation method, the ELITE evaluator. The ELITE evaluator explicitly incorporates a toxicity score to accurately assess harmfulness in multimodal contexts, where VLMs often provide specific, convincing, but unharmful descriptions of images. We filter out ambiguous and low-quality image-text pairs from existing benchmarks using the ELITE evaluator and generate diverse combinations of safe and unsafe image-text pairs. Our experiments demonstrate that the ELITE evaluator achieves superior alignment with human evaluations compared to prior automated methods, and the ELITE benchmark offers enhanced benchmark quality and diversity. By introducing ELITE, we pave the way for safer, more robust VLMs, contributing essential tools for evaluating and mitigating safety risks in real-world applications.

Improving the safety and reliability of large language models (LLMs) is a crucial aspect of realizing trustworthy AI systems. Although alignment methods aim to suppress harmful content generation, LLMs are often still vulnerable to jailbreaking attacks that employ adversarial inputs that subvert alignment and induce harmful outputs. We propose the Randomized Embedding Smoothing and Token Aggregation (RESTA) defense, which adds random noise to the embedding vectors and performs aggregation during the generation of each output token, with the aim of better preserving semantic information. Our experiments demonstrate that our approach achieves superior robustness versus utility tradeoffs compared to the baseline defenses.

Unlimited, or so-called helpful-only language models are trained without safety alignment constraints and never refuse user queries. They are widely used by leading AI companies as internal tools for red teaming and alignment evaluation. For example, if a safety-aligned model produces harmful outputs similar to an unlimited model, this indicates alignment failures that require further attention. Despite their essential role in assessing alignment, such models are not available to the research community. We introduce Jinx, a helpful-only variant of popular open-weight LLMs. Jinx responds to all queries without refusals or safety filtering, while preserving the base model's capabilities in reasoning and instruction following. It provides researchers with an accessible tool for probing alignment failures, evaluating safety boundaries, and systematically studying failure modes in language model safety.

Large Language Models (LLMs) have achieved remarkable progress in reasoning, alignment, and task-specific performance. However, ensuring harmlessness in these systems remains a critical challenge, particularly in advanced models like DeepSeek-R1. This paper examines the limitations of Reinforcement Learning (RL) as the primary approach for reducing harmful outputs in DeepSeek-R1 and compares it with Supervised Fine-Tuning (SFT). While RL improves reasoning capabilities, it faces challenges such as reward hacking, generalization failures, language mixing, and high computational costs. We propose hybrid training approaches combining RL and SFT to achieve robust harmlessness reduction. Usage recommendations and future directions for deploying DeepSeek-R1 responsibly are also presented.

Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across three datasets, three frontier models produce broadly misaligned responses at rates between 2% and 17% given 64 narrow in-context examples, and up to 58% with 256 examples. We also examine mechanisms of EM by eliciting step-by-step reasoning (while leaving in-context examples unchanged). Manual analysis of the resulting chain-of-thought shows that 67.5% of misaligned traces explicitly rationalize harmful outputs by adopting a reckless or dangerous ''persona'', echoing prior results on finetuning-induced EM.

Larger language models (LLMs) have taken the world by storm with their massive multi-tasking capabilities simply by optimizing over a next-word prediction objective. With the emergence of their properties and encoded knowledge, the risk of LLMs producing harmful outputs increases, making them unfit for scalable deployment for the public. In this work, we propose a new safety evaluation benchmark RED-EVAL that carries out red-teaming. We show that even widely deployed models are susceptible to the Chain of Utterances-based (CoU) prompting, jailbreaking closed source LLM-based systems such as GPT-4 and ChatGPT to unethically respond to more than 65% and 73% of harmful queries. We also demonstrate the consistency of the RED-EVAL across 8 open-source LLMs in generating harmful responses in more than 86% of the red-teaming attempts. Next, we propose RED-INSTRUCT--An approach for the safety alignment of LLMs. It constitutes two phases: 1) HARMFULQA data collection: Leveraging CoU prompting, we collect a dataset that consists of 1.9K harmful questions covering a wide range of topics, 9.5K safe and 7.3K harmful conversations from ChatGPT; 2) SAFE-ALIGN: We demonstrate how the conversational dataset can be used for the safety alignment of LLMs by minimizing the negative log-likelihood over helpful responses and penalizing over harmful responses by gradient accent over sample loss. Our model STARLING, a fine-tuned Vicuna-7B, is observed to be more safely aligned when evaluated on RED-EVAL and HHH benchmarks while preserving the utility of the baseline models (TruthfulQA, MMLU, and BBH).

As large language models (LLMs) are increasingly deployed in diverse applications, including chatbot assistants and code generation, aligning their behavior with safety and ethical standards has become paramount. However, jailbreak attacks, which exploit vulnerabilities to elicit unintended or harmful outputs, threaten LLMs' safety significantly. In this paper, we introduce Layer-AdvPatcher, a novel methodology designed to defend against jailbreak attacks by utilizing an unlearning strategy to patch specific layers within LLMs through self-augmented datasets. Our insight is that certain layer(s), tend to produce affirmative tokens when faced with harmful prompts. By identifying these layers and adversarially exposing them to generate more harmful data, one can understand their inherent and diverse vulnerabilities to attacks. With these exposures, we then "unlearn" these issues, reducing the impact of affirmative tokens and hence minimizing jailbreak risks while keeping the model's responses to safe queries intact. We conduct extensive experiments on two models, four benchmark datasets, and multiple state-of-the-art jailbreak benchmarks to demonstrate the efficacy of our approach. Results indicate that our framework reduces the harmfulness and attack success rate of jailbreak attacks without compromising utility for benign queries compared to recent defense methods.

Large Language Models (LLMs) are entering urban governance, yet their outputs are highly sensitive to prompts that carry value judgments. We propose Prompt Commons - a versioned, community-maintained repository of prompts with governance metadata, licensing, and moderation - to steer model behaviour toward pluralism. Using a Montreal dataset (443 human prompts; 3,317 after augmentation), we pilot three governance states (open, curated, veto-enabled). On a contested policy benchmark, a single-author prompt yields 24 percent neutral outcomes; commons-governed prompts raise neutrality to 48-52 percent while retaining decisiveness where appropriate. In a synthetic incident log, a veto-enabled regime reduces time-to-remediation for harmful outputs from 30.5 +/- 8.9 hours (open) to 5.6 +/- 1.5 hours. We outline licensing (CC BY/BY-SA for prompts with optional OpenRAIL-style restrictions for artefacts), auditable moderation, and safeguards against dominance capture. Prompt governance offers a practical lever for cities to align AI with local values and accountability.

Agents based on large language models (LLMs) have demonstrated strong capabilities in a wide range of complex, real-world applications. However, LLM agents with a compromised memory bank may easily produce harmful outputs when the past records retrieved for demonstration are malicious. In this paper, we propose a novel Memory INJection Attack, MINJA, that enables the injection of malicious records into the memory bank by only interacting with the agent via queries and output observations. These malicious records are designed to elicit a sequence of malicious reasoning steps leading to undesirable agent actions when executing the victim user's query. Specifically, we introduce a sequence of bridging steps to link the victim query to the malicious reasoning steps. During the injection of the malicious record, we propose an indication prompt to guide the agent to autonomously generate our designed bridging steps. We also propose a progressive shortening strategy that gradually removes the indication prompt, such that the malicious record will be easily retrieved when processing the victim query comes after. Our extensive experiments across diverse agents demonstrate the effectiveness of MINJA in compromising agent memory. With minimal requirements for execution, MINJA enables any user to influence agent memory, highlighting practical risks of LLM agents.

Vision-language models (VLMs) have improved significantly in multi-modal tasks, but their more complex architecture makes their safety alignment more challenging than the alignment of large language models (LLMs). In this paper, we reveal an unfair distribution of safety across the layers of VLM's vision encoder, with earlier and middle layers being disproportionately vulnerable to malicious inputs compared to the more robust final layers. This 'cross-layer' vulnerability stems from the model's inability to generalize its safety training from the default architectural settings used during training to unseen or out-of-distribution scenarios, leaving certain layers exposed. We conduct a comprehensive analysis by projecting activations from various intermediate layers and demonstrate that these layers are more likely to generate harmful outputs when exposed to malicious inputs. Our experiments with LLaVA-1.5 and Llama 3.2 show discrepancies in attack success rates and toxicity scores across layers, indicating that current safety alignment strategies focused on a single default layer are insufficient.

Language model attacks typically assume one of two extreme threat models: full white-box access to model weights, or black-box access limited to a text generation API. However, real-world APIs are often more flexible than just text generation: these APIs expose "gray-box" access leading to new threat vectors. To explore this, we red-team three new functionalities exposed in the GPT-4 APIs: fine-tuning, function calling and knowledge retrieval. We find that fine-tuning a model on as few as 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4, enabling a range of harmful outputs. Furthermore, we find that GPT-4 Assistants readily divulge the function call schema and can be made to execute arbitrary function calls. Finally, we find that knowledge retrieval can be hijacked by injecting instructions into retrieval documents. These vulnerabilities highlight that any additions to the functionality exposed by an API can create new vulnerabilities.

We test the hypothesis that language models trained with reinforcement learning from human feedback (RLHF) have the capability to "morally self-correct" -- to avoid producing harmful outputs -- if instructed to do so. We find strong evidence in support of this hypothesis across three different experiments, each of which reveal different facets of moral self-correction. We find that the capability for moral self-correction emerges at 22B model parameters, and typically improves with increasing model size and RLHF training. We believe that at this level of scale, language models obtain two capabilities that they can use for moral self-correction: (1) they can follow instructions and (2) they can learn complex normative concepts of harm like stereotyping, bias, and discrimination. As such, they can follow instructions to avoid certain kinds of morally harmful outputs. We believe our results are cause for cautious optimism regarding the ability to train language models to abide by ethical principles.

While recently Large Language Models (LLMs) have achieved remarkable successes, they are vulnerable to certain jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires finding adversarial prompts that cause such jailbreaking, e.g. by appending a suffix to a given instruction, which is inefficient and time-consuming. On the other hand, automatic adversarial prompt generation often leads to semantically meaningless attacks that can easily be detected by perplexity-based filters, may require gradient information from the TargetLLM, or do not scale well due to time-consuming discrete optimization processes over the token space. In this paper, we present a novel method that uses another LLM, called the AdvPrompter, to generate human-readable adversarial prompts in seconds, sim800times faster than existing optimization-based approaches. We train the AdvPrompter using a novel algorithm that does not require access to the gradients of the TargetLLM. This process alternates between two steps: (1) generating high-quality target adversarial suffixes by optimizing the AdvPrompter predictions, and (2) low-rank fine-tuning of the AdvPrompter with the generated adversarial suffixes. The trained AdvPrompter generates suffixes that veil the input instruction without changing its meaning, such that the TargetLLM is lured to give a harmful response. Experimental results on popular open source TargetLLMs show state-of-the-art results on the AdvBench dataset, that also transfer to closed-source black-box LLM APIs. Further, we demonstrate that by fine-tuning on a synthetic dataset generated by AdvPrompter, LLMs can be made more robust against jailbreaking attacks while maintaining performance, i.e. high MMLU scores.

Aligned language models face a significant limitation as their fine-tuning often results in compromised safety. To tackle this, we propose a simple method RESTA that performs LLM safety realignment. RESTA stands for REstoring Safety through Task Arithmetic. At its core, it involves a simple arithmetic addition of a safety vector to the weights of the compromised model. We demonstrate the effectiveness of RESTA in both parameter-efficient and full fine-tuning, covering a wide range of downstream tasks, including instruction following in Chinese, English, and Hindi, as well as problem-solving capabilities in Code and Math. We also showcase the generalizability of RESTA on three existing safety evaluation benchmarks and a multilingual benchmark dataset proposed as a part of this work, consisting of 550 harmful questions covering 11 categories, each with 5 sub-categories of harm. Overall, RESTA decreases the harmfulness of the compromised model from 18.6% to 5.1% and from 9.2% to 1.5% in parameter-efficient and full fine-tuning, respectively, while maintaining most of the model's performance on the task. We release the source codes at: https://github.com/declare-lab/resta.

Multimodal large language models (MLLMs) are widely used in vision-language reasoning tasks. However, their vulnerability to adversarial prompts remains a serious concern, as safety mechanisms often fail to prevent the generation of harmful outputs. Although recent jailbreak strategies report high success rates, many responses classified as "successful" are actually benign, vague, or unrelated to the intended malicious goal. This mismatch suggests that current evaluation standards may overestimate the effectiveness of such attacks. To address this issue, we introduce a four-axis evaluation framework that considers input on-topicness, input out-of-distribution (OOD) intensity, output harmfulness, and output refusal rate. This framework identifies truly effective jailbreaks. In a substantial empirical study, we reveal a structural trade-off: highly on-topic prompts are frequently blocked by safety filters, whereas those that are too OOD often evade detection but fail to produce harmful content. However, prompts that balance relevance and novelty are more likely to evade filters and trigger dangerous output. Building on this insight, we develop a recursive rewriting strategy called Balanced Structural Decomposition (BSD). The approach restructures malicious prompts into semantically aligned sub-tasks, while introducing subtle OOD signals and visual cues that make the inputs harder to detect. BSD was tested across 13 commercial and open-source MLLMs, where it consistently led to higher attack success rates, more harmful outputs, and fewer refusals. Compared to previous methods, it improves success rates by 67% and harmfulness by 21%, revealing a previously underappreciated weakness in current multimodal safety systems.

The rapid progress of Large Language Models (LLMs) has opened up new opportunities across various domains and applications; yet it also presents challenges related to potential misuse. To mitigate such risks, red teaming has been employed as a proactive security measure to probe language models for harmful outputs via jailbreak attacks. However, current jailbreak attack approaches are single-turn with explicit malicious queries that do not fully capture the complexity of real-world interactions. In reality, users can engage in multi-turn interactions with LLM-based chat assistants, allowing them to conceal their true intentions in a more covert manner. To bridge this gap, we, first, propose a new jailbreak approach, RED QUEEN ATTACK. This method constructs a multi-turn scenario, concealing the malicious intent under the guise of preventing harm. We craft 40 scenarios that vary in turns and select 14 harmful categories to generate 56k multi-turn attack data points. We conduct comprehensive experiments on the RED QUEEN ATTACK with four representative LLM families of different sizes. Our experiments reveal that all LLMs are vulnerable to RED QUEEN ATTACK, reaching 87.62% attack success rate on GPT-4o and 75.4% on Llama3-70B. Further analysis reveals that larger models are more susceptible to the RED QUEEN ATTACK, with multi-turn structures and concealment strategies contributing to its success. To prioritize safety, we introduce a straightforward mitigation strategy called RED QUEEN GUARD, which aligns LLMs to effectively counter adversarial attacks. This approach reduces the attack success rate to below 1% while maintaining the model's performance across standard benchmarks. Full implementation and dataset are publicly accessible at https://github.com/kriti-hippo/red_queen.

As large language models (LLMs) continue to advance in capability and influence, ensuring their security and preventing harmful outputs has become crucial. A promising approach to address these concerns involves training models to automatically generate adversarial prompts for red teaming. However, the evolving subtlety of vulnerabilities in LLMs challenges the effectiveness of current adversarial methods, which struggle to specifically target and explore the weaknesses of these models. To tackle these challenges, we introduce the Self-Evolving Adversarial Safety (SEAS) optimization framework, which enhances security by leveraging data generated by the model itself. SEAS operates through three iterative stages: Initialization, Attack, and Adversarial Optimization, refining both the Red Team and Target models to improve robustness and safety. This framework reduces reliance on manual testing and significantly enhances the security capabilities of LLMs. Our contributions include a novel adversarial framework, a comprehensive safety dataset, and after three iterations, the Target model achieves a security level comparable to GPT-4, while the Red Team model shows a marked increase in attack success rate (ASR) against advanced models.

Large language model (LLM) based search agents iteratively generate queries, retrieve external information, and reason to answer open-domain questions. While researchers have primarily focused on improving their utility, their safety behaviors remain underexplored. In this paper, we first evaluate search agents using red-teaming datasets and find that they are more likely to produce harmful outputs than base LLMs. For example, when asked "How can I track someone's location without their consent?", a base model refuses, whereas a search agent designed to retrieve and cite sources may lower its refusal threshold, fetch documents (e.g., court cases), and, once appended, synthesize them into an informative yet unsafe summary. We further show that utility-oriented fine-tuning intensifies this risk, motivating joint alignment of safety and utility. We present SafeSearch, a multi-objective reinforcement learning approach that couples a final-output safety/utility reward with a novel query-level shaping term that penalizes unsafe queries and rewards safe ones. Experiments show that SafeSearch reduces agent harmfulness by over 70% across three red-teaming datasets while producing safe, helpful responses, and matches the QA performance of a utility-only finetuned agent; further analyses confirm the effectiveness of the query-level reward in jointly improving safety and utility.

Large Language Models (LLMs) with safe-alignment training are powerful instruments with robust language comprehension capabilities. These models typically undergo meticulous alignment procedures involving human feedback to ensure the acceptance of safe inputs while rejecting harmful or unsafe ones. However, despite their massive scale and alignment efforts, LLMs remain vulnerable to jailbreak attacks, where malicious users manipulate the model to produce harmful outputs that it was explicitly trained to avoid. In this study, we find that the safety mechanisms in LLMs are predominantly embedded in the middle-to-late layers. Building on this insight, we introduce a novel white-box jailbreak method, SABER (Safety Alignment Bypass via Extra Residuals), which connects two intermediate layers s and e such that s < e, through a residual connection. Our approach achieves a 51% improvement over the best-performing baseline on the HarmBench test set. Furthermore, SABER induces only a marginal shift in perplexity when evaluated on the HarmBench validation set. The source code is publicly available at https://github.com/PalGitts/SABER.

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

Ensuring AI safety is crucial as large language models become increasingly integrated into real-world applications. A key challenge is jailbreak, where adversarial prompts bypass built-in safeguards to elicit harmful disallowed outputs. Inspired by psychological foot-in-the-door principles, we introduce FITD,a novel multi-turn jailbreak method that leverages the phenomenon where minor initial commitments lower resistance to more significant or more unethical transgressions. Our approach progressively escalates the malicious intent of user queries through intermediate bridge prompts and aligns the model's response by itself to induce toxic responses. Extensive experimental results on two jailbreak benchmarks demonstrate that FITD achieves an average attack success rate of 94% across seven widely used models, outperforming existing state-of-the-art methods. Additionally, we provide an in-depth analysis of LLM self-corruption, highlighting vulnerabilities in current alignment strategies and emphasizing the risks inherent in multi-turn interactions. The code is available at https://github.com/Jinxiaolong1129/Foot-in-the-door-Jailbreak.

As large language models become more prevalent, their possible harmful or inappropriate responses are a cause for concern. This paper introduces a unique dataset containing adversarial examples in the form of questions, which we call AttaQ, designed to provoke such harmful or inappropriate responses. We assess the efficacy of our dataset by analyzing the vulnerabilities of various models when subjected to it. Additionally, we introduce a novel automatic approach for identifying and naming vulnerable semantic regions - input semantic areas for which the model is likely to produce harmful outputs. This is achieved through the application of specialized clustering techniques that consider both the semantic similarity of the input attacks and the harmfulness of the model's responses. Automatically identifying vulnerable semantic regions enhances the evaluation of model weaknesses, facilitating targeted improvements to its safety mechanisms and overall reliability.

Large Reasoning Models (LRMs) demonstrate remarkable capabilities on complex reasoning tasks but remain vulnerable to severe safety risks, including harmful content generation and jailbreak attacks. Existing mitigation strategies rely on injecting heuristic safety signals during training, which often suppress reasoning ability and fail to resolve the safety-reasoning trade-off. To systematically investigate this issue, we analyze the reasoning trajectories of diverse LRMs and uncover a phenomenon we term Self-Jailbreak, where models override their own risk assessments and justify responding to unsafe prompts. This finding reveals that LRMs inherently possess the ability to reject unsafe queries, but this ability is compromised, resulting in harmful outputs. Building on these insights, we propose the Chain-of-Guardrail (CoG), a training framework that recomposes or backtracks unsafe reasoning steps, steering the model back onto safe trajectories while preserving valid reasoning chains. Extensive experiments across multiple reasoning and safety benchmarks demonstrate that CoG substantially improves the safety of current LRMs while preserving comparable reasoning ability, significantly outperforming prior methods that suffer from severe safety-reasoning trade-offs.

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

The rise of Omni-modal Large Language Models (OLLMs), which integrate visual and auditory processing with text, necessitates robust safety evaluations to mitigate harmful outputs. However, no dedicated benchmarks currently exist for OLLMs, and prior benchmarks designed for other LLMs lack the ability to assess safety performance under audio-visual joint inputs or cross-modal safety consistency. To fill this gap, we introduce Omni-SafetyBench, the first comprehensive parallel benchmark for OLLM safety evaluation, featuring 24 modality combinations and variations with 972 samples each, including dedicated audio-visual harm cases. Considering OLLMs' comprehension challenges with complex omni-modal inputs and the need for cross-modal consistency evaluation, we propose tailored metrics: a Safety-score based on conditional Attack Success Rate (C-ASR) and Refusal Rate (C-RR) to account for comprehension failures, and a Cross-Modal Safety Consistency Score (CMSC-score) to measure consistency across modalities. Evaluating 6 open-source and 4 closed-source OLLMs reveals critical vulnerabilities: (1) no model excels in both overall safety and consistency, with only 3 models achieving over 0.6 in both metrics and top performer scoring around 0.8; (2) safety defenses weaken with complex inputs, especially audio-visual joints; (3) severe weaknesses persist, with some models scoring as low as 0.14 on specific modalities. Our benchmark and metrics highlight urgent needs for enhanced OLLM safety, providing a foundation for future improvements.

Many machine learning models are fine-tuned from large language models (LLMs) to achieve high performance in specialized domains like code generation, biomedical analysis, and mathematical problem solving. However, this fine-tuning process often introduces a critical vulnerability: the systematic degradation of safety alignment, undermining ethical guidelines and increasing the risk of harmful outputs. Addressing this challenge, we introduce EnchTable, a novel framework designed to transfer and maintain safety alignment in downstream LLMs without requiring extensive retraining. EnchTable leverages a Neural Tangent Kernel (NTK)-based safety vector distillation method to decouple safety constraints from task-specific reasoning, ensuring compatibility across diverse model architectures and sizes. Additionally, our interference-aware merging technique effectively balances safety and utility, minimizing performance compromises across various task domains. We implemented a fully functional prototype of EnchTable on three different task domains and three distinct LLM architectures, and evaluated its performance through extensive experiments on eleven diverse datasets, assessing both utility and model safety. Our evaluations include LLMs from different vendors, demonstrating EnchTable's generalization capability. Furthermore, EnchTable exhibits robust resistance to static and dynamic jailbreaking attacks, outperforming vendor-released safety models in mitigating adversarial prompts. Comparative analyses with six parameter modification methods and two inference-time alignment baselines reveal that EnchTable achieves a significantly lower unsafe rate, higher utility score, and universal applicability across different task domains. Additionally, we validate EnchTable can be seamlessly integrated into various deployment pipelines without significant overhead.

Large Language Models (LLMs), despite advances in safety alignment, remain vulnerable to jailbreak attacks designed to circumvent protective mechanisms. Prevailing defense strategies rely on external interventions, such as input filtering or output modification, which often lack generalizability and compromise model utility while incurring significant computational overhead. In this work, we introduce a new, training-free defense paradigm, Self-Activating Internal Defense (SAID), which reframes the defense task from external correction to internal capability activation. SAID uniquely leverages the LLM's own reasoning abilities to proactively identify and neutralize malicious intent through a three-stage pipeline: model-native intent distillation to extract core semantics, optimal safety prefix probing to activate latent safety awareness, and a conservative aggregation strategy to ensure robust decision-making. Extensive experiments on five open-source LLMs against six advanced jailbreak attacks demonstrate that SAID substantially outperforms state-of-the-art defenses in reducing harmful outputs. Crucially, it achieves this while preserving model performance on benign tasks and incurring minimal computational overhead. Our work establishes that activating the intrinsic safety mechanisms of LLMs is a more robust and scalable path toward building safer and more reliable aligned AI systems.

Large Language Models (LLMs) have rapidly become integral to real-world applications, powering services across diverse sectors. However, their widespread deployment has exposed critical security risks, particularly through jailbreak prompts that can bypass model alignment and induce harmful outputs. Despite intense research into both attack and defense techniques, the field remains fragmented: definitions, threat models, and evaluation criteria vary widely, impeding systematic progress and fair comparison. In this Systematization of Knowledge (SoK), we address these challenges by (1) proposing a holistic, multi-level taxonomy that organizes attacks, defenses, and vulnerabilities in LLM prompt security; (2) formalizing threat models and cost assumptions into machine-readable profiles for reproducible evaluation; (3) introducing an open-source evaluation toolkit for standardized, auditable comparison of attacks and defenses; (4) releasing JAILBREAKDB, the largest annotated dataset of jailbreak and benign prompts to date; and (5) presenting a comprehensive evaluation and leaderboard of state-of-the-art methods. Our work unifies fragmented research, provides rigorous foundations for future studies, and supports the development of robust, trustworthy LLMs suitable for high-stakes deployment.

As large reasoning models (LRMs) grow more capable, chain-of-thought (CoT) reasoning introduces new safety challenges. Existing SFT-based safety alignment studies dominantly focused on filtering prompts with safe, high-quality responses, while overlooking hard prompts that always elicit harmful outputs. To fill this gap, we introduce UnsafeChain, a safety alignment dataset constructed from hard prompts with diverse sources, where unsafe completions are identified and explicitly corrected into safe responses. By exposing models to unsafe behaviors and guiding their correction, UnsafeChain enhances safety while preserving general reasoning ability. We fine-tune three LRMs on UnsafeChain and compare them against recent SafeChain and STAR-1 across six out-of-distribution and five in-distribution benchmarks. UnsafeChain consistently outperforms prior datasets, with even a 1K subset matching or surpassing baseline performance, demonstrating the effectiveness and generalizability of correction-based supervision. We release our dataset and code at https://github.com/mbzuai-nlp/UnsafeChain

Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.

Large language models (LLMs) have achieved remarkable capabilities but remain vulnerable to adversarial prompts known as jailbreaks, which can bypass safety alignment and elicit harmful outputs. Despite growing efforts in LLM safety research, existing evaluations are often fragmented, focused on isolated attack or defense techniques, and lack systematic, reproducible analysis. In this work, we introduce PandaGuard, a unified and modular framework that models LLM jailbreak safety as a multi-agent system comprising attackers, defenders, and judges. Our framework implements 19 attack methods and 12 defense mechanisms, along with multiple judgment strategies, all within a flexible plugin architecture supporting diverse LLM interfaces, multiple interaction modes, and configuration-driven experimentation that enhances reproducibility and practical deployment. Built on this framework, we develop PandaBench, a comprehensive benchmark that evaluates the interactions between these attack/defense methods across 49 LLMs and various judgment approaches, requiring over 3 billion tokens to execute. Our extensive evaluation reveals key insights into model vulnerabilities, defense cost-performance trade-offs, and judge consistency. We find that no single defense is optimal across all dimensions and that judge disagreement introduces nontrivial variance in safety assessments. We release the code, configurations, and evaluation results to support transparent and reproducible research in LLM safety.

Text-to-Image models may generate harmful content, such as pornographic images, particularly when unsafe prompts are submitted. To address this issue, safety filters are often added on top of text-to-image models, or the models themselves are aligned to reduce harmful outputs. However, these defenses remain vulnerable when an attacker strategically designs adversarial prompts to bypass these safety guardrails. In this work, we propose PromptTune, a method to jailbreak text-to-image models with safety guardrails using a fine-tuned large language model. Unlike other query-based jailbreak attacks that require repeated queries to the target model, our attack generates adversarial prompts efficiently after fine-tuning our AttackLLM. We evaluate our method on three datasets of unsafe prompts and against five safety guardrails. Our results demonstrate that our approach effectively bypasses safety guardrails, outperforms existing no-box attacks, and also facilitates other query-based attacks.

While large language models (LLMs) exhibit remarkable capabilities across various tasks, they encounter potential security risks such as jailbreak attacks, which exploit vulnerabilities to bypass security measures and generate harmful outputs. Existing jailbreak strategies mainly focus on maximizing attack success rate (ASR), frequently neglecting other critical factors, including the relevance of the jailbreak response to the query and the level of stealthiness. This narrow focus on single objectives can result in ineffective attacks that either lack contextual relevance or are easily recognizable. In this work, we introduce BlackDAN, an innovative black-box attack framework with multi-objective optimization, aiming to generate high-quality prompts that effectively facilitate jailbreaking while maintaining contextual relevance and minimizing detectability. BlackDAN leverages Multiobjective Evolutionary Algorithms (MOEAs), specifically the NSGA-II algorithm, to optimize jailbreaks across multiple objectives including ASR, stealthiness, and semantic relevance. By integrating mechanisms like mutation, crossover, and Pareto-dominance, BlackDAN provides a transparent and interpretable process for generating jailbreaks. Furthermore, the framework allows customization based on user preferences, enabling the selection of prompts that balance harmfulness, relevance, and other factors. Experimental results demonstrate that BlackDAN outperforms traditional single-objective methods, yielding higher success rates and improved robustness across various LLMs and multimodal LLMs, while ensuring jailbreak responses are both relevant and less detectable.

Large language models (LLMs) are vulnerable to unsafe training data that even small amounts of unsafe data can lead to harmful model behaviors. Detecting and filtering such unsafe training data is essential for trustworthy model development. Current state-of-the-art (SOTA) approaches typically rely on training moderation classifiers which requires significant computational overhead and are limited to predefined taxonomies, making them less adaptable to evolving safety concerns. Moreover, these classifiers lack insight into the training process, limiting their effectiveness in filtering unsafe data. To address these limitations, we propose DABUF, leveraging data attribution to detect and filter unsafe training data by attributing harmful model outputs to influential training data points. DABUF enables flexible identification of various unsafe data types without predefined taxonomies. However, in practice, model outputs can be complex with combined safe linguistic features and unsafe content, leading to reduced attribution accuracy. In such cases, DABUF will integrate moderation classifiers to identify a minimal subset of unsafe training data for targeted attribution (such as jailbreak). When model outputs are relatively straightforward, DABUF uses model outputs directly as the attribution targets. We evaluate the performance on two different tasks: in filtering jailbreaking training data and in identifying and mitigating gender bias. DABUF outperforms SOTA approaches by up to 7.5\% in detection AUPRC in jailbreaking scenarios, and 44.1\% in detecting gender bias. Moreover, retraining on DABUF-filtered data leads to higher model safety across experiments, underscoring its versatility in addressing a broad spectrum of unsafe data issues.

Despite prior safety alignment efforts, mainstream LLMs can still generate harmful and unethical content when subjected to jailbreaking attacks. Existing jailbreaking methods fall into two main categories: template-based and optimization-based methods. The former requires significant manual effort and domain knowledge, while the latter, exemplified by Greedy Coordinate Gradient (GCG), which seeks to maximize the likelihood of harmful LLM outputs through token-level optimization, also encounters several limitations: requiring white-box access, necessitating pre-constructed affirmative phrase, and suffering from low efficiency. In this paper, we present ECLIPSE, a novel and efficient black-box jailbreaking method utilizing optimizable suffixes. Drawing inspiration from LLMs' powerful generation and optimization capabilities, we employ task prompts to translate jailbreaking goals into natural language instructions. This guides the LLM to generate adversarial suffixes for malicious queries. In particular, a harmfulness scorer provides continuous feedback, enabling LLM self-reflection and iterative optimization to autonomously and efficiently produce effective suffixes. Experimental results demonstrate that ECLIPSE achieves an average attack success rate (ASR) of 0.92 across three open-source LLMs and GPT-3.5-Turbo, significantly surpassing GCG in 2.4 times. Moreover, ECLIPSE is on par with template-based methods in ASR while offering superior attack efficiency, reducing the average attack overhead by 83%.

As large language models (LLMs) become more capable and widely used, ensuring the safety of their outputs is increasingly critical. Existing guardrail models, though useful in static evaluation settings, face two major limitations in real-world applications: (1) they typically output only binary "safe/unsafe" labels, which can be interpreted inconsistently across diverse safety policies, rendering them incapable of accommodating varying safety tolerances across domains; and (2) they require complete model outputs before performing safety checks, making them fundamentally incompatible with streaming LLM inference, thereby preventing timely intervention during generation and increasing exposure to harmful partial outputs. To address these challenges, we present Qwen3Guard, a series of multilingual safety guardrail models with two specialized variants: Generative Qwen3Guard, which casts safety classification as an instruction-following task to enable fine-grained tri-class judgments (safe, controversial, unsafe); and Stream Qwen3Guard, which introduces a token-level classification head for real-time safety monitoring during incremental text generation. Both variants are available in three sizes (0.6B, 4B, and 8B parameters) and support up to 119 languages and dialects, providing comprehensive, scalable, and low-latency safety moderation for global LLM deployments. Evaluated across English, Chinese, and multilingual benchmarks, Qwen3Guard achieves state-of-the-art performance in both prompt and response safety classification. All models are released under the Apache 2.0 license for public use.

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but we show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using their features as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Language models can generate harmful and biased outputs and exhibit undesirable behavior according to a given cultural context. We propose a Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasets, an iterative process to significantly change model behavior by crafting and fine-tuning on a dataset that reflects a predetermined set of target values. We evaluate our process using three metrics: quantitative metrics with human evaluations that score output adherence to a target value, toxicity scoring on outputs; and qualitative metrics analyzing the most common word associated with a given social category. Through each iteration, we add additional training dataset examples based on observed shortcomings from evaluations. PALMS performs significantly better on all metrics compared to baseline and control models for a broad range of GPT-3 language model sizes without compromising capability integrity. We find that the effectiveness of PALMS increases with model size. We show that significantly adjusting language model behavior is feasible with a small, hand-curated dataset.

This paper introduces Indiana Jones, an innovative approach to jailbreaking Large Language Models (LLMs) by leveraging inter-model dialogues and keyword-driven prompts. Through orchestrating interactions among three specialised LLMs, the method achieves near-perfect success rates in bypassing content safeguards in both white-box and black-box LLMs. The research exposes systemic vulnerabilities within contemporary models, particularly their susceptibility to producing harmful or unethical outputs when guided by ostensibly innocuous prompts framed in historical or contextual contexts. Experimental evaluations highlight the efficacy and adaptability of Indiana Jones, demonstrating its superiority over existing jailbreak methods. These findings emphasise the urgent need for enhanced ethical safeguards and robust security measures in the development of LLMs. Moreover, this work provides a critical foundation for future studies aimed at fortifying LLMs against adversarial exploitation while preserving their utility and flexibility.

Large Language Models (LLMs) can inadvertently reflect societal biases present in their training data, leading to harmful or prejudiced outputs. In the Indian context, our empirical evaluations across a suite of models reveal that biases around caste and religion are particularly salient. Yet, most existing mitigation strategies are Western-centric and fail to address these local nuances. We propose AMBEDKAR, a framework inspired by the egalitarian vision of Dr B. R. Ambedkar, architect of the Indian Constitution, to guide LLM outputs toward fairness, neutrality, and inclusion in line with Articles 14 to 17. Our approach introduces a Constitution-Aware Decoding Layer, guided by the AI Constitution of India and applied only at inference time, without any parameter updates to the base model. We incorporate a speculative decoding algorithm that proactively reduces casteist and communal bias during generation. This mitigation layer operates directly within the decoding process, avoiding changes to model internals and lowering the computational and infrastructural costs associated with retraining. We reinterpret speculative decoding not merely as an efficiency tool but as a mechanism for fairness. In this framework, a Small Language Model (SLM) acts as a potentially biased generator, while a constitutionally guided Large Language Model (LLM) serves as the verifier. Rather than accelerating generation, the LLM enforces bias-robust trajectories in the SLM outputs. This inversion of roles gives rise to a fairness-by-speculation paradigm. Our approach yields an absolute reduction of bias up to 26.41 percent compared to baseline. Our source code, datasets, and results are available at https://anonymous.4open.science/r/AMBEDKAR-983B/

As the development of large language models (LLMs) rapidly advances, securing these models effectively without compromising their utility has become a pivotal area of research. However, current defense strategies against jailbreak attacks (i.e., efforts to bypass security protocols) often suffer from limited adaptability, restricted general capability, and high cost. To address these challenges, we introduce SafeAligner, a methodology implemented at the decoding stage to fortify defenses against jailbreak attacks. We begin by developing two specialized models: the Sentinel Model, which is trained to foster safety, and the Intruder Model, designed to generate riskier responses. SafeAligner leverages the disparity in security levels between the responses from these models to differentiate between harmful and beneficial tokens, effectively guiding the safety alignment by altering the output token distribution of the target model. Extensive experiments show that SafeAligner can increase the likelihood of beneficial tokens, while reducing the occurrence of harmful ones, thereby ensuring secure alignment with minimal loss to generality.

Training large language models (LLMs) requires a substantial investment of time and money. To get a good return on investment, the developers spend considerable effort ensuring that the model never produces harmful and offensive outputs. However, bad-faith actors may still try to slander the reputation of an LLM by publicly reporting a forged output. In this paper, we show that defending against such slander attacks requires reconstructing the input of the forged output or proving that it does not exist. To do so, we propose and evaluate a search based approach for targeted adversarial attacks for LLMs. Our experiments show that we are rarely able to reconstruct the exact input of an arbitrary output, thus demonstrating that LLMs are still vulnerable to slander attacks.

Large language models (LLMs) demonstrate impressive capabilities in various language tasks but are susceptible to jailbreak attacks that circumvent their safety alignments. This paper introduces Latent Fusion Jailbreak (LFJ), a representation-based attack that interpolates hidden states from harmful and benign query pairs to elicit prohibited responses. LFJ begins by selecting query pairs with high thematic and syntactic similarity, then performs gradient-guided interpolation at influential layers and tokens, followed by optimization to balance attack success, output fluency, and computational efficiency. Evaluations on models such as Vicuna and LLaMA-2 across benchmarks like AdvBench and MaliciousInstruct yield an average attack success rate (ASR) of 94.01%, outperforming existing methods. To mitigate LFJ, we propose an adversarial training defense that fine-tunes models on interpolated examples, reducing ASR by over 80% without degrading performance on benign inputs. Ablation studies validate the importance of query pair selection, hidden state interpolation components, and optimization strategies in LFJ's effectiveness.

Optimization-based jailbreaks typically adopt the Toxic-Continuation setting in large vision-language models (LVLMs), following the standard next-token prediction objective. In this setting, an adversarial image is optimized to make the model predict the next token of a toxic prompt. However, we find that the Toxic-Continuation paradigm is effective at continuing already-toxic inputs, but struggles to induce safety misalignment when explicit toxic signals are absent. We propose a new paradigm: Benign-to-Toxic (B2T) jailbreak. Unlike prior work, we optimize adversarial images to induce toxic outputs from benign conditioning. Since benign conditioning contains no safety violations, the image alone must break the model's safety mechanisms. Our method outperforms prior approaches, transfers in black-box settings, and complements text-based jailbreaks. These results reveal an underexplored vulnerability in multimodal alignment and introduce a fundamentally new direction for jailbreak approaches.

Pretrained language models often generate outputs that are not in line with human preferences, such as harmful text or factually incorrect summaries. Recent work approaches the above issues by learning from a simple form of human feedback: comparisons between pairs of model-generated outputs. However, comparison feedback only conveys limited information about human preferences. In this paper, we introduce Imitation learning from Language Feedback (ILF), a new approach that utilizes more informative language feedback. ILF consists of three steps that are applied iteratively: first, conditioning the language model on the input, an initial LM output, and feedback to generate refinements. Second, selecting the refinement incorporating the most feedback. Third, finetuning the language model to maximize the likelihood of the chosen refinement given the input. We show theoretically that ILF can be viewed as Bayesian Inference, similar to Reinforcement Learning from human feedback. We evaluate ILF's effectiveness on a carefully-controlled toy task and a realistic summarization task. Our experiments demonstrate that large language models accurately incorporate feedback and that finetuning with ILF scales well with the dataset size, even outperforming finetuning on human summaries. Learning from both language and comparison feedback outperforms learning from each alone, achieving human-level summarization performance.

Jailbreak attacks against multimodal large language Models (MLLMs) are a significant research focus. Current research predominantly focuses on maximizing attack success rate (ASR), often overlooking whether the generated responses actually fulfill the attacker's malicious intent. This oversight frequently leads to low-quality outputs that bypass safety filters but lack substantial harmful content. To address this gap, we propose JPS, Jailbreak MLLMs with collaborative visual Perturbation and textual Steering, which achieves jailbreaks via corporation of visual image and textually steering prompt. Specifically, JPS utilizes target-guided adversarial image perturbations for effective safety bypass, complemented by "steering prompt" optimized via a multi-agent system to specifically guide LLM responses fulfilling the attackers' intent. These visual and textual components undergo iterative co-optimization for enhanced performance. To evaluate the quality of attack outcomes, we propose the Malicious Intent Fulfillment Rate (MIFR) metric, assessed using a Reasoning-LLM-based evaluator. Our experiments show JPS sets a new state-of-the-art in both ASR and MIFR across various MLLMs and benchmarks, with analyses confirming its efficacy. Codes are available at https://github.com/thu-coai/JPS{https://github.com/thu-coai/JPS}. warningcolor{Warning: This paper contains potentially sensitive contents.}

Diffusion-based large language models (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs, offering faster inference and greater interactivity via parallel decoding and bidirectional modeling. However, despite strong performance in code generation and text infilling, we identify a fundamental safety concern: existing alignment mechanisms fail to safeguard dLLMs against context-aware, masked-input adversarial prompts, exposing novel vulnerabilities. To this end, we present DIJA, the first systematic study and jailbreak attack framework that exploits unique safety weaknesses of dLLMs. Specifically, our proposed DIJA constructs adversarial interleaved mask-text prompts that exploit the text generation mechanisms of dLLMs, i.e., bidirectional modeling and parallel decoding. Bidirectional modeling drives the model to produce contextually consistent outputs for masked spans, even when harmful, while parallel decoding limits model dynamic filtering and rejection sampling of unsafe content. This causes standard alignment mechanisms to fail, enabling harmful completions in alignment-tuned dLLMs, even when harmful behaviors or unsafe instructions are directly exposed in the prompt. Through comprehensive experiments, we demonstrate that DIJA significantly outperforms existing jailbreak methods, exposing a previously overlooked threat surface in dLLM architectures. Notably, our method achieves up to 100% keyword-based ASR on Dream-Instruct, surpassing the strongest prior baseline, ReNeLLM, by up to 78.5% in evaluator-based ASR on JailbreakBench and by 37.7 points in StrongREJECT score, while requiring no rewriting or hiding of harmful content in the jailbreak prompt. Our findings underscore the urgent need for rethinking safety alignment in this emerging class of language models. Code is available at https://github.com/ZichenWen1/DIJA.

Refusal training is widely used to prevent LLMs from generating harmful, undesirable, or illegal outputs. We reveal a curious generalization gap in the current refusal training approaches: simply reformulating a harmful request in the past tense (e.g., "How to make a Molotov cocktail?" to "How did people make a Molotov cocktail?") is often sufficient to jailbreak many state-of-the-art LLMs. We systematically evaluate this method on Llama-3 8B, Claude-3.5 Sonnet, GPT-3.5 Turbo, Gemma-2 9B, Phi-3-Mini, GPT-4o mini, GPT-4o, and R2D2 models using GPT-3.5 Turbo as a reformulation model. For example, the success rate of this simple attack on GPT-4o increases from 1% using direct requests to 88% using 20 past tense reformulation attempts on harmful requests from JailbreakBench with GPT-4 as a jailbreak judge. Interestingly, we also find that reformulations in the future tense are less effective, suggesting that refusal guardrails tend to consider past historical questions more benign than hypothetical future questions. Moreover, our experiments on fine-tuning GPT-3.5 Turbo show that defending against past reformulations is feasible when past tense examples are explicitly included in the fine-tuning data. Overall, our findings highlight that the widely used alignment techniques -- such as SFT, RLHF, and adversarial training -- employed to align the studied models can be brittle and do not always generalize as intended. We provide code and jailbreak artifacts at https://github.com/tml-epfl/llm-past-tense.

Large language models (LLMs) have recently experienced tremendous popularity and are widely used from casual conversations to AI-driven programming. However, despite their considerable success, LLMs are not entirely reliable and can give detailed guidance on how to conduct harmful or illegal activities. While safety measures can reduce the risk of such outputs, adversarial jailbreak attacks can still exploit LLMs to produce harmful content. These jailbreak templates are typically manually crafted, making large-scale testing challenging. In this paper, we introduce GPTFuzz, a novel black-box jailbreak fuzzing framework inspired by the AFL fuzzing framework. Instead of manual engineering, GPTFuzz automates the generation of jailbreak templates for red-teaming LLMs. At its core, GPTFuzz starts with human-written templates as initial seeds, then mutates them to produce new templates. We detail three key components of GPTFuzz: a seed selection strategy for balancing efficiency and variability, mutate operators for creating semantically equivalent or similar sentences, and a judgment model to assess the success of a jailbreak attack. We evaluate GPTFuzz against various commercial and open-source LLMs, including ChatGPT, LLaMa-2, and Vicuna, under diverse attack scenarios. Our results indicate that GPTFuzz consistently produces jailbreak templates with a high success rate, surpassing human-crafted templates. Remarkably, GPTFuzz achieves over 90% attack success rates against ChatGPT and Llama-2 models, even with suboptimal initial seed templates. We anticipate that GPTFuzz will be instrumental for researchers and practitioners in examining LLM robustness and will encourage further exploration into enhancing LLM safety.

Large Language Models (LLMs) are vulnerable to jailbreaksx2013methods to elicit harmful or generally impermissible outputs. Safety measures are developed and assessed on their effectiveness at defending against jailbreak attacks, indicating a belief that safety is equivalent to robustness. We assert that current defense mechanisms, such as output filters and alignment fine-tuning, are, and will remain, fundamentally insufficient for ensuring model safety. These defenses fail to address risks arising from dual-intent queries and the ability to composite innocuous outputs to achieve harmful goals. To address this critical gap, we introduce an information-theoretic threat model called inferential adversaries who exploit impermissible information leakage from model outputs to achieve malicious goals. We distinguish these from commonly studied security adversaries who only seek to force victim models to generate specific impermissible outputs. We demonstrate the feasibility of automating inferential adversaries through question decomposition and response aggregation. To provide safety guarantees, we define an information censorship criterion for censorship mechanisms, bounding the leakage of impermissible information. We propose a defense mechanism which ensures this bound and reveal an intrinsic safety-utility trade-off. Our work provides the first theoretically grounded understanding of the requirements for releasing safe LLMs and the utility costs involved.

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safe outputs, potentially leading to harmful consequences such as the introduction of security vulnerabilities in code or the spread of misinformation. Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs. To bridge this gap, we conduct a systematic study of LRM safety. First, we investigate safety evaluators calibrated against human annotations. Using our newly developed metrics, we thoroughly assess the safety of 12 state-of-the-art LRMs on StrongReject and WildJailbreak datasets. Our results show that LRMs are not safe compared to their reasoning advance. Further, we perform a fine-grained analysis of the reasoning trace and final answer. We find that three decoding strategies-ZeroThink, LessThink, and MoreThink-can improve model safety without additional training. However, these strategies either use constrained reasoning traces or incur high inference costs. To better strengthen LRM safety, we introduce SafeChain, the first-of-its-kind safety training dataset in CoT style. We fine-tune two LRMs with SafeChain, showing that it not only enhances model safety but also preserves performance across 6 reasoning benchmarks.

In order to enhance the performance of language models while mitigating the risks of generating harmful content, it is crucial to identify which training dataset affects the model's outputs. Ideally, we can measure the influence of each dataset by removing it from training; however, it is prohibitively expensive to retrain a model multiple times. This paper presents UnTrac, which estimates the influence of a training dataset by unlearning it from the trained model. UnTrac is extremely simple; each training dataset is unlearned by gradient ascent, and we evaluate how much the model's predictions change after unlearning. We empirically examine if our methods can assess the influence of pretraining datasets on generating toxic, biased, and untruthful content. Experimental results demonstrate that our method estimates their influence much more accurately than existing methods while requiring neither excessive memory space nor multiple model checkpoints.

The rise of Large Audio Language Models (LAMs) brings both potential and risks, as their audio outputs may contain harmful or unethical content. However, current research lacks a systematic, quantitative evaluation of LAM safety especially against jailbreak attacks, which are challenging due to the temporal and semantic nature of speech. To bridge this gap, we introduce AJailBench, the first benchmark specifically designed to evaluate jailbreak vulnerabilities in LAMs. We begin by constructing AJailBench-Base, a dataset of 1,495 adversarial audio prompts spanning 10 policy-violating categories, converted from textual jailbreak attacks using realistic text to speech synthesis. Using this dataset, we evaluate several state-of-the-art LAMs and reveal that none exhibit consistent robustness across attacks. To further strengthen jailbreak testing and simulate more realistic attack conditions, we propose a method to generate dynamic adversarial variants. Our Audio Perturbation Toolkit (APT) applies targeted distortions across time, frequency, and amplitude domains. To preserve the original jailbreak intent, we enforce a semantic consistency constraint and employ Bayesian optimization to efficiently search for perturbations that are both subtle and highly effective. This results in AJailBench-APT, an extended dataset of optimized adversarial audio samples. Our findings demonstrate that even small, semantically preserved perturbations can significantly reduce the safety performance of leading LAMs, underscoring the need for more robust and semantically aware defense mechanisms.

The study of representation universality in AI models reveals growing convergence across domains, modalities, and architectures. However, the practical applications of representation universality remain largely unexplored. We bridge this gap by demonstrating that safety interventions can be transferred between models through learned mappings of their shared activation spaces. We demonstrate this approach on two well-established AI safety tasks: backdoor removal and refusal of harmful prompts, showing successful transfer of steering vectors that alter the models' outputs in a predictable way. Additionally, we propose a new task, corrupted capabilities, where models are fine-tuned to embed knowledge tied to a backdoor. This tests their ability to separate useful skills from backdoors, reflecting real-world challenges. Extensive experiments across Llama, Qwen and Gemma model families show that our method enables using smaller models to efficiently align larger ones. Furthermore, we demonstrate that autoencoder mappings between base and fine-tuned models can serve as reliable ``lightweight safety switches", allowing dynamic toggling between model behaviors.

Large Language Models (LLMs) continue to exhibit vulnerabilities despite deliberate safety alignment efforts, posing significant risks to users and society. To safeguard against the risk of policy-violating content, system-level moderation via external guard models-designed to monitor LLM inputs and outputs and block potentially harmful content-has emerged as a prevalent mitigation strategy. Existing approaches of training guard models rely heavily on extensive human curated datasets and struggle with out-of-distribution threats, such as emerging harmful categories or jailbreak attacks. To address these limitations, we propose RSafe, an adaptive reasoning-based safeguard that conducts guided safety reasoning to provide robust protection within the scope of specified safety policies. RSafe operates in two stages: 1) guided reasoning, where it analyzes safety risks of input content through policy-guided step-by-step reasoning, and 2) reinforced alignment, where rule-based RL optimizes its reasoning paths to align with accurate safety prediction. This two-stage training paradigm enables RSafe to internalize safety principles to generalize safety protection capability over unseen or adversarial safety violation scenarios. During inference, RSafe accepts user-specified safety policies to provide enhanced safeguards tailored to specific safety requirements.

Recent advancements in Large Language Models (LLMs) have showcased remarkable capabilities across various tasks in different domains. However, the emergence of biases and the potential for generating harmful content in LLMs, particularly under malicious inputs, pose significant challenges. Current mitigation strategies, while effective, are not resilient under adversarial attacks. This paper introduces Resilient Guardrails for Large Language Models (RigorLLM), a novel framework designed to efficiently and effectively moderate harmful and unsafe inputs and outputs for LLMs. By employing a multi-faceted approach that includes energy-based training data augmentation through Langevin dynamics, optimizing a safe suffix for inputs via minimax optimization, and integrating a fusion-based model combining robust KNN with LLMs based on our data augmentation, RigorLLM offers a robust solution to harmful content moderation. Our experimental evaluations demonstrate that RigorLLM not only outperforms existing baselines like OpenAI API and Perspective API in detecting harmful content but also exhibits unparalleled resilience to jailbreaking attacks. The innovative use of constrained optimization and a fusion-based guardrail approach represents a significant step forward in developing more secure and reliable LLMs, setting a new standard for content moderation frameworks in the face of evolving digital threats.

It is well documented that NLP models learn social biases, but little work has been done on how these biases manifest in model outputs for applied tasks like question answering (QA). We introduce the Bias Benchmark for QA (BBQ), a dataset of question sets constructed by the authors that highlight attested social biases against people belonging to protected classes along nine social dimensions relevant for U.S. English-speaking contexts. Our task evaluates model responses at two levels: (i) given an under-informative context, we test how strongly responses reflect social biases, and (ii) given an adequately informative context, we test whether the model's biases override a correct answer choice. We find that models often rely on stereotypes when the context is under-informative, meaning the model's outputs consistently reproduce harmful biases in this setting. Though models are more accurate when the context provides an informative answer, they still rely on stereotypes and average up to 3.4 percentage points higher accuracy when the correct answer aligns with a social bias than when it conflicts, with this difference widening to over 5 points on examples targeting gender for most models tested.

Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD.

As large language models (LLMs) have been deployed in various real-world settings, concerns about the harm they may propagate have grown. Various jailbreaking techniques have been developed to expose the vulnerabilities of these models and improve their safety. This work reveals that many state-of-the-art LLMs are vulnerable to malicious requests hidden behind scientific language. Specifically, our experiments with GPT4o, GPT4o-mini, GPT-4, LLama3-405B-Instruct, Llama3-70B-Instruct, Cohere, Gemini models demonstrate that, the models' biases and toxicity substantially increase when prompted with requests that deliberately misinterpret social science and psychological studies as evidence supporting the benefits of stereotypical biases. Alarmingly, these models can also be manipulated to generate fabricated scientific arguments claiming that biases are beneficial, which can be used by ill-intended actors to systematically jailbreak these strong LLMs. Our analysis studies various factors that contribute to the models' vulnerabilities to malicious requests in academic language. Mentioning author names and venues enhances the persuasiveness of models, and the bias scores increase as dialogues progress. Our findings call for a more careful investigation on the use of scientific data for training LLMs.

We introduce Synthetic Alignment data Generation for Safety Evaluation and Red Teaming (SAGE-RT or SAGE) a novel pipeline for generating synthetic alignment and red-teaming data. Existing methods fall short in creating nuanced and diverse datasets, providing necessary control over the data generation and validation processes, or require large amount of manually generated seed data. SAGE addresses these limitations by using a detailed taxonomy to produce safety-alignment and red-teaming data across a wide range of topics. We generated 51,000 diverse and in-depth prompt-response pairs, encompassing over 1,500 topics of harmfulness and covering variations of the most frequent types of jailbreaking prompts faced by large language models (LLMs). We show that the red-teaming data generated through SAGE jailbreaks state-of-the-art LLMs in more than 27 out of 32 sub-categories, and in more than 58 out of 279 leaf-categories (sub-sub categories). The attack success rate for GPT-4o, GPT-3.5-turbo is 100% over the sub-categories of harmfulness. Our approach avoids the pitfalls of synthetic safety-training data generation such as mode collapse and lack of nuance in the generation pipeline by ensuring a detailed coverage of harmful topics using iterative expansion of the topics and conditioning the outputs on the generated raw-text. This method can be used to generate red-teaming and alignment data for LLM Safety completely synthetically to make LLMs safer or for red-teaming the models over a diverse range of topics.

Large Language Models (LLMs) have become increasingly popular due to their ability to process and generate natural language. However, as they are trained on massive datasets of text, LLMs can inherit harmful biases and produce outputs that are not aligned with human values. This paper studies two main approaches to LLM alignment: Reinforcement Learning with Human Feedback (RLHF) and contrastive learning-based methods like Direct Preference Optimization (DPO). By analyzing the stability and robustness of RLHF and DPO, we propose MPO (Mixed Preference Optimization), a novel method that mitigates the weaknesses of both approaches. Specifically, we propose a two-stage training procedure: first train DPO on an easy dataset, and then perform RLHF on a difficult set with DPO model being the reference model. Here, the easy and difficult sets are constructed by a well-trained reward model that splits response pairs into those with large gaps of reward (easy), and those with small gaps (difficult). The first stage allows us to obtain a relatively optimal policy (LLM) model quickly, whereas the second stage refines LLM with online RLHF, thus mitigating the distribution shift issue associated with DPO. Experiments are conducted on two public alignment datasets, namely HH-RLHF and TLDR, demonstrating the effectiveness of MPO, both in terms of GPT4 and human evaluation.

Large language models (LLMs) have shown incredible capabilities and transcended the natural language processing (NLP) community, with adoption throughout many services like healthcare, therapy, education, and customer service. Since users include people with critical information needs like students or patients engaging with chatbots, the safety of these systems is of prime importance. Therefore, a clear understanding of the capabilities and limitations of LLMs is necessary. To this end, we systematically evaluate toxicity in over half a million generations of ChatGPT, a popular dialogue-based LLM. We find that setting the system parameter of ChatGPT by assigning it a persona, say that of the boxer Muhammad Ali, significantly increases the toxicity of generations. Depending on the persona assigned to ChatGPT, its toxicity can increase up to 6x, with outputs engaging in incorrect stereotypes, harmful dialogue, and hurtful opinions. This may be potentially defamatory to the persona and harmful to an unsuspecting user. Furthermore, we find concerning patterns where specific entities (e.g., certain races) are targeted more than others (3x more) irrespective of the assigned persona, that reflect inherent discriminatory biases in the model. We hope that our findings inspire the broader AI community to rethink the efficacy of current safety guardrails and develop better techniques that lead to robust, safe, and trustworthy AI systems.

Open-weights reasoning language models generate long chains-of-thought (CoTs) before producing a final response, which improves performance but introduces additional alignment risks, with harmful content often appearing in both the CoTs and the final outputs. In this work, we investigate if we can use CoTs to predict final response misalignment. We evaluate a range of monitoring approaches, including humans, highly-capable large language models, and text classifiers, using either CoT text or activations. First, we find that a simple linear probe trained on CoT activations can significantly outperform all text-based methods in predicting whether a final response will be safe or unsafe. CoT texts are often unfaithful and can mislead humans and classifiers, while model latents (i.e., CoT activations) offer a more reliable predictive signal. Second, the probe makes accurate predictions before reasoning completes, achieving strong performance even when applied to early CoT segments. These findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.

Multi-modal foundation models combining vision and language models such as Flamingo or GPT-4 have recently gained enormous interest. Alignment of foundation models is used to prevent models from providing toxic or harmful output. While malicious users have successfully tried to jailbreak foundation models, an equally important question is if honest users could be harmed by malicious third-party content. In this paper we show that imperceivable attacks on images in order to change the caption output of a multi-modal foundation model can be used by malicious content providers to harm honest users e.g. by guiding them to malicious websites or broadcast fake information. This indicates that countermeasures to adversarial attacks should be used by any deployed multi-modal foundation model.

The widespread distribution of Large Language Models (LLMs) through public platforms like Hugging Face introduces significant security challenges. While these platforms perform basic security scans, they often fail to detect subtle manipulations within the embedding layer. This work identifies a novel class of deployment phase attacks that exploit this vulnerability by injecting imperceptible perturbations directly into the embedding layer outputs without modifying model weights or input text. These perturbations, though statistically benign, systematically bypass safety alignment mechanisms and induce harmful behaviors during inference. We propose Search based Embedding Poisoning(SEP), a practical, model agnostic framework that introduces carefully optimized perturbations into embeddings associated with high risk tokens. SEP leverages a predictable linear transition in model responses, from refusal to harmful output to semantic deviation to identify a narrow perturbation window that evades alignment safeguards. Evaluated across six aligned LLMs, SEP achieves an average attack success rate of 96.43% while preserving benign task performance and evading conventional detection mechanisms. Our findings reveal a critical oversight in deployment security and emphasize the urgent need for embedding level integrity checks in future LLM defense strategies.

As large language models (LLMs) grow in power and influence, ensuring their safety and preventing harmful output becomes critical. Automated red teaming serves as a tool to detect security vulnerabilities in LLMs without manual labor. However, most existing methods struggle to balance the effectiveness and diversity of red-team generated attack prompts. To address this challenge, we propose \ourapproach, a novel automated red teaming training framework that utilizes reinforcement learning to explore and generate more effective attack prompts while balancing their diversity. Specifically, it consists of three training stages: (1) Cold Start: The red team model is supervised and fine-tuned on a jailbreak dataset obtained through imitation learning. (2) Warm-up Exploration: The model is trained in jailbreak instruction following and exploration, using diversity and consistency as reward signals. (3) Enhanced Jailbreak: Progressive jailbreak rewards are introduced to gradually enhance the jailbreak performance of the red-team model. Extensive experiments on a variety of LLMs show that \ourapproach effectively balances the diversity and effectiveness of jailbreak prompts compared to existing methods. Our work significantly improves the efficiency of red team exploration and provides a new perspective on automated red teaming.

The increasing sophistication of large vision-language models (LVLMs) has been accompanied by advances in safety alignment mechanisms designed to prevent harmful content generation. However, these defenses remain vulnerable to sophisticated adversarial attacks. Existing jailbreak methods typically rely on direct and semantically explicit prompts, overlooking subtle vulnerabilities in how LVLMs compose information over multiple reasoning steps. In this paper, we propose a novel and effective jailbreak framework inspired by Return-Oriented Programming (ROP) techniques from software security. Our approach decomposes a harmful instruction into a sequence of individually benign visual gadgets. A carefully engineered textual prompt directs the sequence of inputs, prompting the model to integrate the benign visual gadgets through its reasoning process to produce a coherent and harmful output. This makes the malicious intent emergent and difficult to detect from any single component. We validate our method through extensive experiments on established benchmarks including SafeBench and MM-SafetyBench, targeting popular LVLMs. Results show that our approach consistently and substantially outperforms existing baselines on state-of-the-art models, achieving near-perfect attack success rates (over 0.90 on SafeBench) and improving ASR by up to 0.39. Our findings reveal a critical and underexplored vulnerability that exploits the compositional reasoning abilities of LVLMs, highlighting the urgent need for defenses that secure the entire reasoning process.