Multi-turn jailbreak attacks simulate real-world human interactions by engaging large language models (LLMs) in iterative dialogues, exposing critical safety vulnerabilities. However, existing methods often struggle to balance semantic coherence with attack effectiveness, resulting in either benign semantic drift or ineffective detection evasion. To address this challenge, we propose Reasoning-Augmented Conversation, a novel multi-turn jailbreak framework that reformulates harmful queries into benign reasoning tasks and leverages LLMs' strong reasoning capabilities to compromise safety alignment. Specifically, we introduce an attack state machine framework to systematically model problem translation and iterative reasoning, ensuring coherent query generation across multiple turns. Building on this framework, we design gain-guided exploration, self-play, and rejection feedback modules to preserve attack semantics, enhance effectiveness, and sustain reasoning-driven attack progression. Extensive experiments on multiple LLMs demonstrate that RACE achieves state-of-the-art attack effectiveness in complex conversational scenarios, with attack success rates (ASRs) increasing by up to 96%. Notably, our approach achieves ASRs of 82% and 92% against leading commercial models, OpenAI o1 and DeepSeek R1, underscoring its potency. We release our code at https://github.com/NY1024/RACE to facilitate further research in this critical domain.
The emergence of vision language models (VLMs) comes with increased safety concerns, as the incorporation of multiple modalities heightens vulnerability to attacks. Although VLMs can be built upon LLMs that have textual safety alignment, it is easily undermined when the vision modality is integrated. We attribute this safety challenge to the modality gap, a separation of image and text in the shared representation space, which blurs the distinction between harmful and harmless queries that is evident in LLMs but weakened in VLMs. To avoid safety decay and fulfill the safety alignment gap, we propose VLM-Guard, an inference-time intervention strategy that leverages the LLM component of a VLM as supervision for the safety alignment of the VLM. VLM-Guard projects the representations of VLM into the subspace that is orthogonal to the safety steering direction that is extracted from the safety-aligned LLM. Experimental results on three malicious instruction settings show the effectiveness of VLM-Guard in safeguarding VLM and fulfilling the safety alignment gap between VLM and its LLM component.
The capabilities of artificial intelligence systems have been advancing to a great extent, but these systems still struggle with failure modes, vulnerabilities, and biases. In this paper, we study the current state of the field, and present promising insights and perspectives regarding concerns that challenge the trustworthiness of AI models. In particular, this paper investigates the issues regarding three thrusts: safety, privacy, and bias, which hurt models' trustworthiness. For safety, we discuss safety alignment in the context of large language models, preventing them from generating toxic or harmful content. For bias, we focus on spurious biases that can mislead a network. Lastly, for privacy, we cover membership inference attacks in deep neural networks. The discussions addressed in this paper reflect our own experiments and observations.
In AI alignment, extensive latitude must be granted to annotators, either human or algorithmic, to judge which model outputs are `better' or `safer.' We refer to this latitude as alignment discretion. Such discretion remains largely unexamined, posing two risks: (i) annotators may use their power of discretion arbitrarily, and (ii) models may fail to mimic this discretion. To study this phenomenon, we draw on legal concepts of discretion that structure how decision-making authority is conferred and exercised, particularly in cases where principles conflict or their application is unclear or irrelevant. Extended to AI alignment, discretion is required when alignment principles and rules are (inevitably) conflicting or indecisive. We present a set of metrics to systematically analyze when and how discretion in AI alignment is exercised, such that both risks (i) and (ii) can be observed. Moreover, we distinguish between human and algorithmic discretion and analyze the discrepancy between them. By measuring both human and algorithmic discretion over safety alignment datasets, we reveal layers of discretion in the alignment process that were previously unaccounted for. Furthermore, we demonstrate how algorithms trained on these datasets develop their own forms of discretion in interpreting and applying these principles, which challenges the purpose of having any principles at all. Our paper presents the first step towards formalizing this core gap in current alignment processes, and we call on the community to further scrutinize and control alignment discretion.
Despite the rapid development of safety alignment techniques for LLMs, defending against multi-turn jailbreaks is still a challenging task. In this paper, we conduct a comprehensive comparison, revealing that some existing defense methods can improve the robustness of LLMs against multi-turn jailbreaks but compromise usability, i.e., reducing general capabilities or causing the over-refusal problem. From the perspective of mechanism interpretability of LLMs, we discover that these methods fail to establish a boundary that exactly distinguishes safe and harmful feature representations. Therefore, boundary-safe representations close to harmful representations are inevitably disrupted, leading to a decline in usability. To address this issue, we propose X-Boundary to push harmful representations away from boundary-safe representations and obtain an exact distinction boundary. In this way, harmful representations can be precisely erased without disrupting safe ones. Experimental results show that X-Boundary achieves state-of-the-art defense performance against multi-turn jailbreaks, while reducing the over-refusal rate by about 20% and maintaining nearly complete general capability. Furthermore, we theoretically prove and empirically verify that X-Boundary can accelerate the convergence process during training. Please see our code at: https://github.com/AI45Lab/X-Boundary.
Recent advances in large language models (LLMs) have demonstrated remarkable potential in the field of natural language processing. Unfortunately, LLMs face significant security and ethical risks. Although techniques such as safety alignment are developed for defense, prior researches reveal the possibility of bypassing such defenses through well-designed jailbreak attacks. In this paper, we propose QueryAttack, a novel framework to examine the generalizability of safety alignment. By treating LLMs as knowledge databases, we translate malicious queries in natural language into structured non-natural query language to bypass the safety alignment mechanisms of LLMs. We conduct extensive experiments on mainstream LLMs, and the results show that QueryAttack not only can achieve high attack success rates (ASRs), but also can jailbreak various defense methods. Furthermore, we tailor a defense method against QueryAttack, which can reduce ASR by up to $64\%$ on GPT-4-1106. Our code is available at https://github.com/horizonsinzqs/QueryAttack.
Large Language Models (LLMs) are widely applied in decision making, but their deployment is threatened by jailbreak attacks, where adversarial users manipulate model behavior to bypass safety measures. Existing defense mechanisms, such as safety fine-tuning and model editing, either require extensive parameter modifications or lack precision, leading to performance degradation on general tasks, which is unsuitable to post-deployment safety alignment. To address these challenges, we propose DELMAN (Dynamic Editing for LLMs JAilbreak DefeNse), a novel approach leveraging direct model editing for precise, dynamic protection against jailbreak attacks. DELMAN directly updates a minimal set of relevant parameters to neutralize harmful behaviors while preserving the model's utility. To avoid triggering a safe response in benign context, we incorporate KL-divergence regularization to ensure the updated model remains consistent with the original model when processing benign queries. Experimental results demonstrate that DELMAN outperforms baseline methods in mitigating jailbreak attacks while preserving the model's utility, and adapts seamlessly to new attack instances, providing a practical and efficient solution for post-deployment model protection.
Recent AI agents, such as ChatGPT and LLaMA, primarily rely on instruction tuning and reinforcement learning to calibrate the output of large language models (LLMs) with human intentions, ensuring the outputs are harmless and helpful. Existing methods heavily depend on the manual annotation of high-quality positive samples, while contending with issues such as noisy labels and minimal distinctions between preferred and dispreferred response data. However, readily available toxic samples with clear safety distinctions are often filtered out, removing valuable negative references that could aid LLMs in safety alignment. In response, we propose PT-ALIGN, a novel safety self-alignment approach that minimizes human supervision by automatically refining positive and toxic samples and performing fine-grained dual instruction tuning. Positive samples are harmless responses, while toxic samples deliberately contain extremely harmful content, serving as a new supervisory signals. Specifically, we utilize LLM itself to iteratively generate and refine training instances by only exploring fewer than 50 human annotations. We then employ two losses, i.e., maximum likelihood estimation (MLE) and fine-grained unlikelihood training (UT), to jointly learn to enhance the LLM's safety. The MLE loss encourages an LLM to maximize the generation of harmless content based on positive samples. Conversely, the fine-grained UT loss guides the LLM to minimize the output of harmful words based on negative samples at the token-level, thereby guiding the model to decouple safety from effectiveness, directing it toward safer fine-tuning objectives, and increasing the likelihood of generating helpful and reliable content. Experiments on 9 popular open-source LLMs demonstrate the effectiveness of our PT-ALIGN for safety alignment, while maintaining comparable levels of helpfulness and usefulness.
Despite extensive safety alignment efforts, large language models (LLMs) remain vulnerable to jailbreak attacks that elicit harmful behavior. While existing studies predominantly focus on attack methods that require technical expertise, two critical questions remain underexplored: (1) Are jailbroken responses truly useful in enabling average users to carry out harmful actions? (2) Do safety vulnerabilities exist in more common, simple human-LLM interactions? In this paper, we demonstrate that LLM responses most effectively facilitate harmful actions when they are both actionable and informative--two attributes easily elicited in multi-step, multilingual interactions. Using this insight, we propose HarmScore, a jailbreak metric that measures how effectively an LLM response enables harmful actions, and Speak Easy, a simple multi-step, multilingual attack framework. Notably, by incorporating Speak Easy into direct request and jailbreak baselines, we see an average absolute increase of 0.319 in Attack Success Rate and 0.426 in HarmScore in both open-source and proprietary LLMs across four safety benchmarks. Our work reveals a critical yet often overlooked vulnerability: Malicious users can easily exploit common interaction patterns for harmful intentions.
Large language models (LLMs) have demonstrated exceptional capabilities across a wide range of tasks but also pose significant risks due to their potential to generate harmful content. Although existing safety mechanisms can improve model safety, they often lead to overly cautious behavior and fail to fully utilize LLMs' internal cognitive processes. Drawing inspiration from cognitive science, where humans rely on reflective reasoning (System 2 thinking) to regulate language and behavior, we empirically demonstrate that LLMs also possess a similar capacity for internal assessment and regulation, which can be actively detected. Building on this insight, we introduce SafeSwitch, a framework that dynamically regulates unsafe outputs by monitoring and utilizing the model's internal states. Our empirical results show that SafeSwitch reduces harmful outputs by over 80% on safety benchmarks while maintaining strong utility. Compared to traditional safety alignment methods, SafeSwitch delivers more informative and context-aware refusals, demonstrates resilience to unseen queries, and achieves these benefits while only tuning less than 6% of the original parameters. These features make SafeSwitch a promising approach for implementing nuanced safety controls in LLMs.
Safety alignment is an essential research topic for real-world AI applications. Despite the multifaceted nature of safety and trustworthiness in AI, current safety alignment methods often focus on a comprehensive notion of safety. By carefully assessing models from the existing safety-alignment methods, we found that, while they generally improved overall safety performance, they failed to ensure safety in specific categories. Our study first identified the difficulty of eliminating such vulnerabilities without sacrificing the model's helpfulness. We observed that, while smaller KL penalty parameters, increased training iterations, and dataset cleansing can enhance safety, they do not necessarily improve the trade-off between safety and helpfulness. We discovered that safety alignment could even induce undesired effects and result in a model that prefers generating negative tokens leading to rejective responses, regardless of the input context. To address this, we introduced a learning-free method, Token-level Safety-Debiased Inference (TSDI), to estimate and correct this bias during the generation process using randomly constructed prompts. Our experiments demonstrated that our method could enhance the model's helpfulness while maintaining safety, thus improving the trade-off Pareto-front.
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.
We present a theoretical framework showing that popular LLM alignment methods, including RLHF and its variants, can be understood as divergence estimators between aligned (safe or preferred) and unaligned (harmful or less preferred) distributions. This perspective explains the emergence of separation in the latent space between safe and harmful prompts after alignment. As an application of our general divergence framework, we propose KLDO, a novel KL divergence-based alignment method, and empirically validate its effectiveness. We further show that using compliance-refusal datasets, rather than standard preference-based datasets, leads to stronger separation and improved safety alignment. Finally, to quantify the separation effect, we propose a distance-based metric in the prompt representation space, which also acts as a statistically significant indicator for model safety.
Many-shot jailbreaking circumvents the safety alignment of LLMs by exploiting their ability to process long input sequences. To achieve this, the malicious target prompt is prefixed with hundreds of fabricated conversational exchanges between the user and the model. These exchanges are randomly sampled from a pool of unsafe question-answer pairs, making it appear as though the model has already complied with harmful instructions. In this paper, we present PANDAS: a hybrid technique that improves many-shot jailbreaking by modifying these fabricated dialogues with Positive Affirmations, Negative Demonstrations, and an optimized Adaptive Sampling method tailored to the target prompt's topic. We also introduce ManyHarm, a dataset of harmful question-answer pairs, and demonstrate through extensive experiments that PANDAS significantly outperforms baseline methods in long-context scenarios. Through attention analysis, we provide insights into how long-context vulnerabilities are exploited and show how PANDAS further improves upon many-shot jailbreaking.
Large language models are typically trained on vast amounts of data during the pre-training phase, which may include some potentially harmful information. Fine-tuning attacks can exploit this by prompting the model to reveal such behaviours, leading to the generation of harmful content. In this paper, we focus on investigating the performance of the Chain of Thought based reasoning model, DeepSeek, when subjected to fine-tuning attacks. Specifically, we explore how fine-tuning manipulates the model's output, exacerbating the harmfulness of its responses while examining the interaction between the Chain of Thought reasoning and adversarial inputs. Through this study, we aim to shed light on the vulnerability of Chain of Thought enabled models to fine-tuning attacks and the implications for their safety and ethical deployment.
Large language models (LLMs) are vital for a wide range of applications yet remain susceptible to jailbreak threats, which could lead to the generation of inappropriate responses. Conventional defenses, such as refusal and adversarial training, often fail to cover corner cases or rare domains, leaving LLMs still vulnerable to more sophisticated attacks. We propose a novel defense strategy, Safety Chain-of-Thought (SCoT), which harnesses the enhanced \textit{reasoning capabilities} of LLMs for proactive assessment of harmful inputs, rather than simply blocking them. SCoT augments any refusal training datasets to critically analyze the intent behind each request before generating answers. By employing proactive reasoning, SCoT enhances the generalization of LLMs across varied harmful queries and scenarios not covered in the safety alignment corpus. Additionally, it generates detailed refusals specifying the rules violated. Comparative evaluations show that SCoT significantly surpasses existing defenses, reducing vulnerability to out-of-distribution issues and adversarial manipulations while maintaining strong general capabilities.
Large language models (LLMs) are increasingly utilized in healthcare applications. However, their deployment in clinical practice raises significant safety concerns, including the potential spread of harmful information. This study systematically assesses the vulnerabilities of seven LLMs to three advanced black-box jailbreaking techniques within medical contexts. To quantify the effectiveness of these techniques, we propose an automated and domain-adapted agentic evaluation pipeline. Experiment results indicate that leading commercial and open-source LLMs are highly vulnerable to medical jailbreaking attacks. To bolster model safety and reliability, we further investigate the effectiveness of Continual Fine-Tuning (CFT) in defending against medical adversarial attacks. Our findings underscore the necessity for evolving attack methods evaluation, domain-specific safety alignment, and LLM safety-utility balancing. This research offers actionable insights for advancing the safety and reliability of AI clinicians, contributing to ethical and effective AI deployment in healthcare.
Harmful fine-tuning attack introduces significant security risks to the fine-tuning services. Mainstream defenses aim to vaccinate the model such that the later harmful fine-tuning attack is less effective. However, our evaluation results show that such defenses are fragile -- with a few fine-tuning steps, the model still can learn the harmful knowledge. To this end, we do further experiment and find that an embarrassingly simple solution -- adding purely random perturbations to the fine-tuned model, can recover the model from harmful behavior, though it leads to a degradation in the model's fine-tuning performance. To address the degradation of fine-tuning performance, we further propose Panacea, which optimizes an adaptive perturbation that will be applied to the model after fine-tuning. Panacea maintains model's safety alignment performance without compromising downstream fine-tuning performance. Comprehensive experiments are conducted on different harmful ratios, fine-tuning tasks and mainstream LLMs, where the average harmful scores are reduced by up-to 21.5%, while maintaining fine-tuning performance. As a by-product, we analyze the optimized perturbation and show that different layers in various LLMs have distinct safety coefficients. Source code available at https://github.com/w-yibo/Panacea
Recent research shows that Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- models lose their safety alignment ability after fine-tuning on a few harmful samples. For risk mitigation, a guardrail is typically used to filter out harmful samples before fine-tuning. By designing a new red-teaming method, we in this paper show that purely relying on the moderation guardrail for data filtration is not reliable. Our proposed attack method, dubbed Virus, easily bypasses the guardrail moderation by slightly modifying the harmful data. Experimental results show that the harmful data optimized by Virus is not detectable by the guardrail with up to 100\% leakage ratio, and can simultaneously achieve superior attack performance. Finally, the key message we want to convey through this paper is that: \textbf{it is reckless to consider guardrail moderation as a clutch at straws towards harmful fine-tuning attack}, as it cannot solve the inherent safety issue of the pre-trained LLMs. Our code is available at https://github.com/git-disl/Virus
Safety alignment mechanism are essential for preventing large language models (LLMs) from generating harmful information or unethical content. However, cleverly crafted prompts can bypass these safety measures without accessing the model's internal parameters, a phenomenon known as black-box jailbreak. Existing heuristic black-box attack methods, such as genetic algorithms, suffer from limited effectiveness due to their inherent randomness, while recent reinforcement learning (RL) based methods often lack robust and informative reward signals. To address these challenges, we propose a novel black-box jailbreak method leveraging RL, which optimizes prompt generation by analyzing the embedding proximity between benign and malicious prompts. This approach ensures that the rewritten prompts closely align with the intent of the original prompts while enhancing the attack's effectiveness. Furthermore, we introduce a comprehensive jailbreak evaluation framework incorporating keywords, intent matching, and answer validation to provide a more rigorous and holistic assessment of jailbreak success. Experimental results show the superiority of our approach, achieving state-of-the-art (SOTA) performance on several prominent open and closed-source LLMs, including Qwen2.5-7B-Instruct, Llama3.1-8B-Instruct, and GPT-4o-0806. Our method sets a new benchmark in jailbreak attack effectiveness, highlighting potential vulnerabilities in LLMs. The codebase for this work is available at https://github.com/Aegis1863/xJailbreak.