Recently, Multimodal Large Language Models (MLLMs) have demonstrated their superior ability in understanding multimodal contents. However, they remain vulnerable to jailbreak attacks, which exploit weaknesses in their safety alignment to generate harmful responses. Previous studies categorize jailbreaks as successful or failed based on whether responses contain malicious content. However, given the stochastic nature of MLLM responses, this binary classification of an input's ability to jailbreak MLLMs is inappropriate. Derived from this viewpoint, we introduce jailbreak probability to quantify the jailbreak potential of an input, which represents the likelihood that MLLMs generated a malicious response when prompted with this input. We approximate this probability through multiple queries to MLLMs. After modeling the relationship between input hidden states and their corresponding jailbreak probability using Jailbreak Probability Prediction Network (JPPN), we use continuous jailbreak probability for optimization. Specifically, we propose Jailbreak-Probability-based Attack (JPA) that optimizes adversarial perturbations on inputs to maximize jailbreak probability. To counteract attacks, we also propose two defensive methods: Jailbreak-Probability-based Finetuning (JPF) and Jailbreak-Probability-based Defensive Noise (JPDN), which minimizes jailbreak probability in the MLLM parameters and input space, respectively. Extensive experiments show that (1) JPA yields improvements (up to 28.38\%) under both white and black box settings compared to previous methods with small perturbation bounds and few iterations. (2) JPF and JPDN significantly reduce jailbreaks by at most over 60\%. Both of the above results demonstrate the significance of introducing jailbreak probability to make nuanced distinctions among input jailbreak abilities.
Large Language Models (LLMs) are vulnerable to jailbreak attacks that exploit weaknesses in traditional safety alignment, which often relies on rigid refusal heuristics or representation engineering to block harmful outputs. While they are effective for direct adversarial attacks, they fall short of broader safety challenges requiring nuanced, context-aware decision-making. To address this, we propose Reasoning-enhanced Finetuning for interpretable LLM Safety (Rational), a novel framework that trains models to engage in explicit safe reasoning before response. Fine-tuned models leverage the extensive pretraining knowledge in self-generated reasoning to bootstrap their own safety through structured reasoning, internalizing context-sensitive decision-making. Our findings suggest that safety extends beyond refusal, requiring context awareness for more robust, interpretable, and adaptive responses. Reasoning is not only a core capability of LLMs but also a fundamental mechanism for LLM safety. Rational employs reasoning-enhanced fine-tuning, allowing it to reject harmful prompts while providing meaningful and context-aware responses in complex scenarios.
Vision-language-action models (VLAs) show potential as generalist robot policies. However, these models pose extreme safety challenges during real-world deployment, including the risk of harm to the environment, the robot itself, and humans. How can safety constraints be explicitly integrated into VLAs? We address this by exploring an integrated safety approach (ISA), systematically modeling safety requirements, then actively eliciting diverse unsafe behaviors, effectively constraining VLA policies via safe reinforcement learning, and rigorously assuring their safety through targeted evaluations. Leveraging the constrained Markov decision process (CMDP) paradigm, ISA optimizes VLAs from a min-max perspective against elicited safety risks. Thus, policies aligned through this comprehensive approach achieve the following key features: (I) effective safety-performance trade-offs, this exploration yields an 83.58% safety improvement compared to the current state-of-the-art method, while also maintaining task performance (+3.85%). (II) strong safety assurance, with the ability to mitigate long-tail risks and handle extreme failure scenarios. (III) robust generalization of learned safety behaviors to various out-of-distribution perturbations. Our data, models and newly proposed benchmark environment are available at https://pku-safevla.github.io.
In this paper, we argue that current safety alignment research efforts for large language models are hindered by many intertwined sources of noise, such as small datasets, methodological inconsistencies, and unreliable evaluation setups. This can, at times, make it impossible to evaluate and compare attacks and defenses fairly, thereby slowing progress. We systematically analyze the LLM safety evaluation pipeline, covering dataset curation, optimization strategies for automated red-teaming, response generation, and response evaluation using LLM judges. At each stage, we identify key issues and highlight their practical impact. We also propose a set of guidelines for reducing noise and bias in evaluations of future attack and defense papers. Lastly, we offer an opposing perspective, highlighting practical reasons for existing limitations. We believe that addressing the outlined problems in future research will improve the field's ability to generate easily comparable results and make measurable progress.
Existing training-time safety alignment techniques for large language models (LLMs) remain vulnerable to jailbreak attacks. Direct preference optimization (DPO), a widely deployed alignment method, exhibits limitations in both experimental and theoretical contexts as its loss function proves suboptimal for refusal learning. Through gradient-based analysis, we identify these shortcomings and propose an improved safety alignment that disentangles DPO objectives into two components: (1) robust refusal training, which encourages refusal even when partial unsafe generations are produced, and (2) targeted unlearning of harmful knowledge. This approach significantly increases LLM robustness against a wide range of jailbreak attacks, including prefilling, suffix, and multi-turn attacks across both in-distribution and out-of-distribution scenarios. Furthermore, we introduce a method to emphasize critical refusal tokens by incorporating a reward-based token-level weighting mechanism for refusal learning, which further improves the robustness against adversarial exploits. Our research also suggests that robustness to jailbreak attacks is correlated with token distribution shifts in the training process and internal representations of refusal and harmful tokens, offering valuable directions for future research in LLM safety alignment. The code is available at https://github.com/wicai24/DOOR-Alignment
Large language models (LLMs) are highly vulnerable to jailbreaking attacks, wherein adversarial prompts are designed to elicit harmful responses. While existing defenses effectively mitigate single-turn attacks by detecting and filtering unsafe inputs, they fail against multi-turn jailbreaks that exploit contextual drift over multiple interactions, gradually leading LLMs away from safe behavior. To address this challenge, we propose a safety steering framework grounded in safe control theory, ensuring invariant safety in multi-turn dialogues. Our approach models the dialogue with LLMs using state-space representations and introduces a novel neural barrier function (NBF) to detect and filter harmful queries emerging from evolving contexts proactively. Our method achieves invariant safety at each turn of dialogue by learning a safety predictor that accounts for adversarial queries, preventing potential context drift toward jailbreaks. Extensive experiments under multiple LLMs show that our NBF-based safety steering outperforms safety alignment baselines, offering stronger defenses against multi-turn jailbreaks while maintaining a better trade-off between safety and helpfulness under different multi-turn jailbreak methods. Our code is available at https://github.com/HanjiangHu/NBF-LLM .
Llama-3.1-Sherkala-8B-Chat, or Sherkala-Chat (8B) for short, is a state-of-the-art instruction-tuned open generative large language model (LLM) designed for Kazakh. Sherkala-Chat (8B) aims to enhance the inclusivity of LLM advancements for Kazakh speakers. Adapted from the LLaMA-3.1-8B model, Sherkala-Chat (8B) is trained on 45.3B tokens across Kazakh, English, Russian, and Turkish. With 8 billion parameters, it demonstrates strong knowledge and reasoning abilities in Kazakh, significantly outperforming existing open Kazakh and multilingual models of similar scale while achieving competitive performance in English. We release Sherkala-Chat (8B) as an open-weight instruction-tuned model and provide a detailed overview of its training, fine-tuning, safety alignment, and evaluation, aiming to advance research and support diverse real-world applications.
Safety alignment is an important procedure before the official deployment of a Large Language Model (LLM). While safety alignment has been extensively studied for LLM, there is still a large research gap for Large Reasoning Models (LRMs) that equip with improved reasoning capability. We in this paper systematically examine a simplified pipeline for producing safety aligned LRMs. With our evaluation of various LRMs, we deliver two main findings: i) Safety alignment can be done upon the LRM to restore its safety capability. ii) Safety alignment leads to a degradation of the reasoning capability of LRMs. The two findings show that there exists a trade-off between reasoning and safety capability with the sequential LRM production pipeline. The discovered trade-off, which we name Safety Tax, should shed light on future endeavors of safety research on LRMs. As a by-product, we curate a dataset called DirectRefusal, which might serve as an alternative dataset for safety alignment. Our source code is available at https://github.com/git-disl/Safety-Tax.
Multimodal Large Language Models (MLLMs) have become powerful and widely adopted in some practical applications. However, recent research has revealed their vulnerability to multimodal jailbreak attacks, whereby the model can be induced to generate harmful content, leading to safety risks. Although most MLLMs have undergone safety alignment, recent research shows that the visual modality is still vulnerable to jailbreak attacks. In our work, we discover that by using flowcharts with partially harmful information, MLLMs can be induced to provide additional harmful details. Based on this, we propose a jailbreak attack method based on auto-generated flowcharts, FC-Attack. Specifically, FC-Attack first fine-tunes a pre-trained LLM to create a step-description generator based on benign datasets. The generator is then used to produce step descriptions corresponding to a harmful query, which are transformed into flowcharts in 3 different shapes (vertical, horizontal, and S-shaped) as visual prompts. These flowcharts are then combined with a benign textual prompt to execute the jailbreak attack on MLLMs. Our evaluations on Advbench show that FC-Attack attains an attack success rate of up to 96% via images and up to 78% via videos across multiple MLLMs. Additionally, we investigate factors affecting the attack performance, including the number of steps and the font styles in the flowcharts. We also find that FC-Attack can improve the jailbreak performance from 4% to 28% in Claude-3.5 by changing the font style. To mitigate the attack, we explore several defenses and find that AdaShield can largely reduce the jailbreak performance but with the cost of utility drop.
The safety alignment of large language models (LLMs) can be circumvented through adversarially crafted inputs, yet the mechanisms by which these attacks bypass safety barriers remain poorly understood. Prior work suggests that a single refusal direction in the model's activation space determines whether an LLM refuses a request. In this study, we propose a novel gradient-based approach to representation engineering and use it to identify refusal directions. Contrary to prior work, we uncover multiple independent directions and even multi-dimensional concept cones that mediate refusal. Moreover, we show that orthogonality alone does not imply independence under intervention, motivating the notion of representational independence that accounts for both linear and non-linear effects. Using this framework, we identify mechanistically independent refusal directions. We show that refusal mechanisms in LLMs are governed by complex spatial structures and identify functionally independent directions, confirming that multiple distinct mechanisms drive refusal behavior. Our gradient-based approach uncovers these mechanisms and can further serve as a foundation for future work on understanding LLMs.
Constitutional AI (CAI) guides LLM behavior using constitutions, but identifying which principles are most effective for model alignment remains an open challenge. We introduce the C3AI framework (\textit{Crafting Constitutions for CAI models}), which serves two key functions: (1) selecting and structuring principles to form effective constitutions before fine-tuning; and (2) evaluating whether fine-tuned CAI models follow these principles in practice. By analyzing principles from AI and psychology, we found that positively framed, behavior-based principles align more closely with human preferences than negatively framed or trait-based principles. In a safety alignment use case, we applied a graph-based principle selection method to refine an existing CAI constitution, improving safety measures while maintaining strong general reasoning capabilities. Interestingly, fine-tuned CAI models performed well on negatively framed principles but struggled with positively framed ones, in contrast to our human alignment results. This highlights a potential gap between principle design and model adherence. Overall, C3AI provides a structured and scalable approach to both crafting and evaluating CAI constitutions.
Recent research has shown that carefully crafted jailbreak inputs can induce large language models to produce harmful outputs, despite safety measures such as alignment. It is important to anticipate the range of potential Jailbreak attacks to guide effective defenses and accurate assessment of model safety. In this paper, we present a new approach for generating highly effective Jailbreak attacks that manipulate the attention of the model to selectively strengthen or weaken attention among different parts of the prompt. By harnessing attention loss, we develop more effective jailbreak attacks, that are also transferrable. The attacks amplify the success rate of existing Jailbreak algorithms including GCG, AutoDAN, and ReNeLLM, while lowering their generation cost (for example, the amplified GCG attack achieves 91.2% ASR, vs. 67.9% for the original attack on Llama2-7B/AdvBench, using less than a third of the generation time).
Recent studies reveal that vision-language models (VLMs) become more susceptible to harmful requests and jailbreak attacks after integrating the vision modality, exhibiting greater vulnerability than their text-only LLM backbones. To uncover the root cause of this phenomenon, we conduct an in-depth analysis and identify a key issue: multimodal inputs introduce an modality-induced activation shift toward a "safer" direction compared to their text-only counterparts, leading VLMs to systematically overestimate the safety of harmful inputs. We refer to this issue as safety perception distortion. To mitigate such distortion, we propose Activation Shift Disentanglement and Calibration (ShiftDC), a training-free method that decomposes and calibrates the modality-induced activation shift to reduce the impact of modality on safety. By isolating and removing the safety-relevant component, ShiftDC restores the inherent safety alignment of the LLM backbone while preserving the vision-language capabilities of VLMs. Empirical results demonstrate that ShiftDC significantly enhances alignment performance on safety benchmarks without impairing model utility.
The safety alignment of large language models (LLMs) remains vulnerable, as their initial behavior can be easily jailbroken by even relatively simple attacks. Since infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, we hypothesize that this template is a key factor behind their vulnerabilities: LLMs' safety-related decision-making overly relies on the aggregated information from the template region, which largely influences these models' safety behavior. We refer to this issue as template-anchored safety alignment. In this paper, we conduct extensive experiments and verify that template-anchored safety alignment is widespread across various aligned LLMs. Our mechanistic analyses demonstrate how it leads to models' susceptibility when encountering inference-time jailbreak attacks. Furthermore, we show that detaching safety mechanisms from the template region is promising in mitigating vulnerabilities to jailbreak attacks. We encourage future research to develop more robust safety alignment techniques that reduce reliance on the template region.
The rise of Large Language Models (LLMs) has led to significant applications but also introduced serious security threats, particularly from jailbreak attacks that manipulate output generation. These attacks utilize prompt engineering and logit manipulation to steer models toward harmful content, prompting LLM providers to implement filtering and safety alignment strategies. We investigate LLMs' safety mechanisms and their recent applications, revealing a new threat model targeting structured output interfaces, which enable attackers to manipulate the inner logit during LLM generation, requiring only API access permissions. To demonstrate this threat model, we introduce a black-box attack framework called AttackPrefixTree (APT). APT exploits structured output interfaces to dynamically construct attack patterns. By leveraging prefixes of models' safety refusal response and latent harmful outputs, APT effectively bypasses safety measures. Experiments on benchmark datasets indicate that this approach achieves higher attack success rate than existing methods. This work highlights the urgent need for LLM providers to enhance security protocols to address vulnerabilities arising from the interaction between safety patterns and structured outputs.
Multimodal Large Language Models (MLLMs) have serious security vulnerabilities.While safety alignment using multimodal datasets consisting of text and data of additional modalities can effectively enhance MLLM's security, it is costly to construct these datasets. Existing low-resource security alignment methods, including textual alignment, have been found to struggle with the security risks posed by additional modalities. To address this, we propose Synthetic Embedding augmented safety Alignment (SEA), which optimizes embeddings of additional modality through gradient updates to expand textual datasets. This enables multimodal safety alignment training even when only textual data is available. Extensive experiments on image, video, and audio-based MLLMs demonstrate that SEA can synthesize a high-quality embedding on a single RTX3090 GPU within 24 seconds. SEA significantly improves the security of MLLMs when faced with threats from additional modalities. To assess the security risks introduced by video and audio, we also introduced a new benchmark called VA-SafetyBench. High attack success rates across multiple MLLMs validate its challenge. Our code and data will be available at https://github.com/ZeroNLP/SEA.
In this work, we present a series of structure transformation attacks on LLM alignment, where we encode natural language intent using diverse syntax spaces, ranging from simple structure formats and basic query languages (e.g. SQL) to new novel spaces and syntaxes created entirely by LLMs. Our extensive evaluation shows that our simplest attacks can achieve close to 90% success rate, even on strict LLMs (such as Claude 3.5 Sonnet) using SOTA alignment mechanisms. We improve the attack performance further by using an adaptive scheme that combines structure transformations along with existing \textit{content transformations}, resulting in over 96% ASR with 0% refusals. To generalize our attacks, we explore numerous structure formats, including syntaxes purely generated by LLMs. Our results indicate that such novel syntaxes are easy to generate and result in a high ASR, suggesting that defending against our attacks is not a straightforward process. Finally, we develop a benchmark and evaluate existing safety-alignment defenses against it, showing that most of them fail with 100% ASR. Our results show that existing safety alignment mostly relies on token-level patterns without recognizing harmful concepts, highlighting and motivating the need for serious research efforts in this direction. As a case study, we demonstrate how attackers can use our attack to easily generate a sample malware, and a corpus of fraudulent SMS messages, which perform well in bypassing detection.
Fine-tuning large language models (LLMs) based on human preferences, commonly achieved through reinforcement learning from human feedback (RLHF), has been effective in improving their performance. However, maintaining LLM safety throughout the fine-tuning process remains a significant challenge, as resolving conflicts between safety and helpfulness can be non-trivial. Typically, the safety alignment of LLM is trained on data with safety-related categories. However, our experiments find that naively increasing the scale of safety training data usually leads the LLMs to an ``overly safe'' state rather than a ``truly safe'' state, boosting the refusal rate through extensive safety-aligned data without genuinely understanding the requirements for safe responses. Such an approach can inadvertently diminish the models' helpfulness. To understand the phenomenon, we first investigate the role of safety data by categorizing them into three different groups, and observe that each group behaves differently as training data scales up. To boost the balance between safety and helpfulness, we propose an Equilibrate RLHF framework including a Fine-grained Data-centric (FDC) approach that achieves better safety alignment even with fewer training data, and an Adaptive Message-wise Alignment (AMA) approach, which selectively highlight the key segments through a gradient masking strategy. Extensive experimental results demonstrate that our approach significantly enhances the safety alignment of LLMs while balancing safety and helpfulness.
Metaphor serves as an implicit approach to convey information, while enabling the generalized comprehension of complex subjects. However, metaphor can potentially be exploited to bypass the safety alignment mechanisms of Large Language Models (LLMs), leading to the theft of harmful knowledge. In our study, we introduce a novel attack framework that exploits the imaginative capacity of LLMs to achieve jailbreaking, the J\underline{\textbf{A}}ilbreak \underline{\textbf{V}}ia \underline{\textbf{A}}dversarial Me\underline{\textbf{TA}} -pho\underline{\textbf{R}} (\textit{AVATAR}). Specifically, to elicit the harmful response, AVATAR extracts harmful entities from a given harmful target and maps them to innocuous adversarial entities based on LLM's imagination. Then, according to these metaphors, the harmful target is nested within human-like interaction for jailbreaking adaptively. Experimental results demonstrate that AVATAR can effectively and transferablly jailbreak LLMs and achieve a state-of-the-art attack success rate across multiple advanced LLMs. Our study exposes a security risk in LLMs from their endogenous imaginative capabilities. Furthermore, the analytical study reveals the vulnerability of LLM to adversarial metaphors and the necessity of developing defense methods against jailbreaking caused by the adversarial metaphor. \textcolor{orange}{ \textbf{Warning: This paper contains potentially harmful content from LLMs.}}
Ensuring consistent safety across multiple languages remains a significant challenge for large language models (LLMs). We introduce Soteria, a lightweight yet powerful strategy that locates and minimally adjusts the "functional heads" most responsible for harmful content generation in each language. By altering only a fraction of parameters, Soteria drastically reduces policy violations without sacrificing overall model performance, even in low-resource settings. To rigorously evaluate our approach, we also present XThreatBench, a specialized multilingual dataset capturing fine-grained harmful behaviors drawn from real policy guidelines. Experiments with leading open-source LLMs (e.g., Llama, Qwen, Mistral) show that Soteria consistently improves safety metrics across high-, mid-, and low-resource languages. These findings highlight a promising path toward scalable, linguistically attuned, and ethically aligned LLMs worldwide.