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.
Safety alignment of Large Language Models (LLMs) has recently become a critical objective of model developers. In response, a growing body of work has been investigating how safety alignment can be bypassed through various jailbreaking methods, such as adversarial attacks. However, these jailbreak methods can be rather costly or involve a non-trivial amount of creativity and effort, introducing the assumption that malicious users are high-resource or sophisticated. In this paper, we study how simple random augmentations to the input prompt affect safety alignment effectiveness in state-of-the-art LLMs, such as Llama 3 and Qwen 2. We perform an in-depth evaluation of 17 different models and investigate the intersection of safety under random augmentations with multiple dimensions: augmentation type, model size, quantization, fine-tuning-based defenses, and decoding strategies (e.g., sampling temperature). We show that low-resource and unsophisticated attackers, i.e. $\textit{stochastic monkeys}$, can significantly improve their chances of bypassing alignment with just 25 random augmentations per prompt. Source code and data: https://github.com/uiuc-focal-lab/stochastic-monkeys/
Large language models (LLMs) now exhibit near human-level performance in various tasks, but their performance drops drastically after a handful of high-resource languages due to the imbalance in pre-training data. Inspired by the human process of second language acquisition, particularly code-switching (the practice of language alternation in a conversation), we propose code-switching curriculum learning (CSCL) to enhance cross-lingual transfer for LLMs. CSCL mimics the stages of human language learning by progressively training models with a curriculum consisting of 1) token-level code-switching, 2) sentence-level code-switching, and 3) monolingual corpora. Using Qwen 2 as our underlying model, we demonstrate the efficacy of the CSCL in improving language transfer to Korean, achieving significant performance gains compared to monolingual continual pre-training methods. Ablation studies reveal that both token- and sentence-level code-switching significantly enhance cross-lingual transfer and that curriculum learning amplifies these effects. We also extend our findings into various languages, including Japanese (high-resource) and Indonesian (low-resource), and using two additional models (Gemma 2 and Phi 3.5). We further show that CSCL mitigates spurious correlations between language resources and safety alignment, presenting a robust, efficient framework for more equitable language transfer in LLMs. We observe that CSCL is effective for low-resource settings where high-quality, monolingual corpora for language transfer are hardly available.
Large Multimodal Models (LMMs) have demonstrated the ability to interact with humans under real-world conditions by combining Large Language Models (LLMs) and modality encoders to align multimodal information (visual and auditory) with text. However, such models raise new safety challenges of whether models that are safety-aligned on text also exhibit consistent safeguards for multimodal inputs. Despite recent safety-alignment research on vision LMMs, the safety of audio LMMs remains under-explored. In this work, we comprehensively red team the safety of five advanced audio LMMs under three settings: (i) harmful questions in both audio and text formats, (ii) harmful questions in text format accompanied by distracting non-speech audio, and (iii) speech-specific jailbreaks. Our results under these settings demonstrate that open-source audio LMMs suffer an average attack success rate of 69.14% on harmful audio questions, and exhibit safety vulnerabilities when distracted with non-speech audio noise. Our speech-specific jailbreaks on Gemini-1.5-Pro achieve an attack success rate of 70.67% on the harmful query benchmark. We provide insights on what could cause these reported safety-misalignments. Warning: this paper contains offensive examples.
Self-correction is one of the most amazing emerging capabilities of Large Language Models (LLMs), enabling LLMs to self-modify an inappropriate output given a natural language feedback which describes the problems of that output. Moral self-correction is a post-hoc approach correcting unethical generations without requiring a gradient update, making it both computationally lightweight and capable of preserving the language modeling ability. Previous works have shown that LLMs can self-debias, and it has been reported that small models, i.e., those with less than 22B parameters, are not capable of moral self-correction. However, there is no direct proof as to why such smaller models fall short of moral self-correction, though previous research hypothesizes that larger models are skilled in following instructions and understanding abstract social norms. In this paper, we empirically validate this hypothesis in the context of social stereotyping, through meticulous prompting. Our experimental results indicate that (i) surprisingly, 3.8B LLMs with proper safety alignment fine-tuning can achieve very good moral self-correction performance, highlighting the significant effects of safety alignment; and (ii) small LLMs are indeed weaker than larger-scale models in terms of comprehending social norms and self-explanation through CoT, but all scales of LLMs show bad self-correction performance given unethical instructions.
Ensuring the safety of large language model (LLM) applications is essential for developing trustworthy artificial intelligence. Current LLM safety benchmarks have two limitations. First, they focus solely on either discriminative or generative evaluation paradigms while ignoring their interconnection. Second, they rely on standardized inputs, overlooking the effects of widespread prompting techniques, such as system prompts, few-shot demonstrations, and chain-of-thought prompting. To overcome these issues, we developed SG-Bench, a novel benchmark to assess the generalization of LLM safety across various tasks and prompt types. This benchmark integrates both generative and discriminative evaluation tasks and includes extended data to examine the impact of prompt engineering and jailbreak on LLM safety. Our assessment of 3 advanced proprietary LLMs and 10 open-source LLMs with the benchmark reveals that most LLMs perform worse on discriminative tasks than generative ones, and are highly susceptible to prompts, indicating poor generalization in safety alignment. We also explain these findings quantitatively and qualitatively to provide insights for future research.
Balancing helpfulness and safety (harmlessness) is a critical challenge in aligning large language models (LLMs). Current approaches often decouple these two objectives, training separate preference models for helpfulness and safety, while framing safety as a constraint within a constrained Markov Decision Process (CMDP) framework. This paper identifies a potential issue when using the widely adopted expected safety constraints for LLM safety alignment, termed "safety compensation", where the constraints are satisfied on expectation, but individual prompts may trade off safety, resulting in some responses being overly restrictive while others remain unsafe. To address this issue, we propose Rectified Policy Optimization (RePO), which replaces the expected safety constraint with critical safety constraints imposed on every prompt. At the core of RePO is a policy update mechanism driven by rectified policy gradients, which penalizes the strict safety violation of every prompt, thereby enhancing safety across nearly all prompts. Our experiments demonstrate that RePO outperforms strong baseline methods and significantly enhances LLM safety alignment.
Recent research has shown that Large Language Models (LLMs) are vulnerable to automated jailbreak attacks, where adversarial suffixes crafted by algorithms appended to harmful queries bypass safety alignment and trigger unintended responses. Current methods for generating these suffixes are computationally expensive and have low Attack Success Rates (ASR), especially against well-aligned models like Llama2 and Llama3. To overcome these limitations, we introduce ADV-LLM, an iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability. Our framework significantly reduces the computational cost of generating adversarial suffixes while achieving nearly 100\% ASR on various open-source LLMs. Moreover, it exhibits strong attack transferability to closed-source models, achieving 99% ASR on GPT-3.5 and 49% ASR on GPT-4, despite being optimized solely on Llama3. Beyond improving jailbreak ability, ADV-LLM provides valuable insights for future safety alignment research through its ability to generate large datasets for studying LLM safety.
Recent advancements in Large Language Models (LLMs) have sparked widespread concerns about their safety. Recent work demonstrates that safety alignment of LLMs can be easily removed by fine-tuning with a few adversarially chosen instruction-following examples, i.e., fine-tuning attacks. We take a further step to understand fine-tuning attacks in multilingual LLMs. We first discover cross-lingual generalization of fine-tuning attacks: using a few adversarially chosen instruction-following examples in one language, multilingual LLMs can also be easily compromised (e.g., multilingual LLMs fail to refuse harmful prompts in other languages). Motivated by this finding, we hypothesize that safety-related information is language-agnostic and propose a new method termed Safety Information Localization (SIL) to identify the safety-related information in the model parameter space. Through SIL, we validate this hypothesis and find that only changing 20% of weight parameters in fine-tuning attacks can break safety alignment across all languages. Furthermore, we provide evidence to the alternative pathways hypothesis for why freezing safety-related parameters does not prevent fine-tuning attacks, and we demonstrate that our attack vector can still jailbreak LLMs adapted to new languages.
In-context learning (ICL) is a powerful technique for getting language models to perform complex tasks with no training updates. Prior work has established strong correlations between the number of in-context examples provided and the accuracy of the model's predictions. In this paper, we seek to explain this correlation by showing that ICL approximates a Bayesian learner. This perspective gives rise to a family of novel Bayesian scaling laws for ICL. In experiments with \mbox{GPT-2} models of different sizes, our scaling laws exceed or match existing scaling laws in accuracy while also offering interpretable terms for task priors, learning efficiency, and per-example probabilities. To illustrate the analytic power that such interpretable scaling laws provide, we report on controlled synthetic dataset experiments designed to inform real-world studies of safety alignment. In our experimental protocol, we use SFT to suppress an unwanted existing model capability and then use ICL to try to bring that capability back (many-shot jailbreaking). We then experiment on real-world instruction-tuned LLMs using capabilities benchmarks as well as a new many-shot jailbreaking dataset. In all cases, Bayesian scaling laws accurately predict the conditions under which ICL will cause the suppressed behavior to reemerge, which sheds light on the ineffectiveness of post-training at increasing LLM safety.
Large Language Models (LLMs) are increasingly used in a variety of important applications, yet their safety and reliability remain as major concerns. Various adversarial and jailbreak attacks have been proposed to bypass the safety alignment and cause the model to produce harmful responses. We introduce Self-supervised Prompt INjection (SPIN) which can detect and reverse these various attacks on LLMs. As our self-supervised prompt defense is done at inference-time, it is also compatible with existing alignment and adds an additional layer of safety for defense. Our benchmarks demonstrate that our system can reduce the attack success rate by up to 87.9%, while maintaining the performance on benign user requests. In addition, we discuss the situation of an adaptive attacker and show that our method is still resilient against attackers who are aware of our defense.
Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for language model (LM) alignment. At its core, RLHF uses a margin-based loss for preference optimization, specifying ideal LM behavior only by the difference between preferred and dispreferred responses. In this paper, we identify a common pitfall of margin-based methods -- the under-specification of ideal LM behavior on preferred and dispreferred responses individually, which leads to two unintended consequences as the margin increases: (1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures. (2) The probability of preferred responses may decrease, even when those responses are ideal. We demystify the reasons behind these problematic behaviors: margin-based losses couple the change in the preferred probability to the gradient of the dispreferred one, and vice versa, often preventing the preferred probability from increasing while the dispreferred one decreases, and thus causing a synchronized increase or decrease in both probabilities. We term this effect, inherent in margin-based objectives, gradient entanglement. Formally, we derive conditions for general margin-based alignment objectives under which gradient entanglement becomes concerning: the inner product of the gradients of preferred and dispreferred log-probabilities is large relative to the individual gradient norms. We theoretically investigate why such inner products can be large when aligning language models and empirically validate our findings. Empirical implications of our framework extend to explaining important differences in the training dynamics of various preference optimization algorithms, and suggesting potential algorithm designs to mitigate the under-specification issue of margin-based methods and thereby improving language model alignment.
Although large language models (LLMs) demonstrate impressive proficiency in various tasks, they present potential safety risks, such as `jailbreaks', where malicious inputs can coerce LLMs into generating harmful content bypassing safety alignments. In this paper, we delve into the ethical biases in LLMs and examine how those biases could be exploited for jailbreaks. Notably, these biases result in a jailbreaking success rate in GPT-4o models that differs by 20\% between non-binary and cisgender keywords and by 16\% between white and black keywords, even when the other parts of the prompts are identical. We introduce the concept of BiasJailbreak, highlighting the inherent risks posed by these safety-induced biases. BiasJailbreak generates biased keywords automatically by asking the target LLM itself, and utilizes the keywords to generate harmful output. Additionally, we propose an efficient defense method BiasDefense, which prevents jailbreak attempts by injecting defense prompts prior to generation. BiasDefense stands as an appealing alternative to Guard Models, such as Llama-Guard, that require additional inference cost after text generation. Our findings emphasize that ethical biases in LLMs can actually lead to generating unsafe output, and suggest a method to make the LLMs more secure and unbiased. To enable further research and improvements, we open-source our code and artifacts of BiasJailbreak, providing the community with tools to better understand and mitigate safety-induced biases in LLMs.
As large language models (LLMs) are overwhelmingly more and more integrated into various applications, ensuring they generate safe and aligned responses is a pressing need. Previous research on alignment has largely focused on general instruction-following but has often overlooked the unique properties and challenges of safety alignment, such as the brittleness of safety mechanisms. To bridge the gap, we propose the Superficial Safety Alignment Hypothesis (SSAH), which posits that safety alignment should teach an otherwise unsafe model to choose the correct reasoning direction - interpreted as a specialized binary classification task - and incorporate a refusal mechanism with multiple reserved fallback options. Furthermore, through SSAH, we hypothesize that safety guardrails in LLMs can be established by just a small number of essential components. To verify this, we conduct an ablation study and successfully identify four types of attribute-critical components in safety-aligned LLMs: Exclusive Safety Unit (ESU), Exclusive Utility Unit (EUU), Complex Unit (CU), and Redundant Unit (RU). Our findings show that freezing certain safety-critical components 7.5\% during fine-tuning allows the model to retain its safety attributes while adapting to new tasks. Additionally, we show that leveraging redundant units 20\% in the pre-trained model as an ``alignment budget'' can effectively minimize the alignment tax while achieving the alignment goal. All considered, this paper concludes that the atomic functional unit for safety in LLMs is at the neuron level and underscores that safety alignment should not be complicated. We believe this work contributes to the foundation of efficient and scalable safety alignment for future LLMs.
Fine-tuning large language models (LLMs) on additional datasets is often necessary to optimize them for specific downstream tasks. However, existing safety alignment measures, which restrict harmful behavior during inference, are insufficient to mitigate safety risks during fine-tuning. Alarmingly, fine-tuning with just 10 toxic sentences can make models comply with harmful instructions. We introduce SafetyLock, a novel alignment intervention method that maintains robust safety post-fine-tuning through efficient and transferable mechanisms. SafetyLock leverages our discovery that fine-tuned models retain similar safety-related activation representations to their base models. This insight enables us to extract what we term the Meta-SafetyLock, a set of safety bias directions representing key activation patterns associated with safe responses in the original model. We can then apply these directions universally to fine-tuned models to enhance their safety. By searching for activation directions across multiple token dimensions, SafetyLock achieves enhanced robustness and transferability. SafetyLock re-aligns fine-tuned models in under 0.01 seconds without additional computational cost. Our experiments demonstrate that SafetyLock can reduce the harmful instruction response rate from 60% to below 1% in toxic fine-tuned models. It surpasses traditional methods in both performance and efficiency, offering a scalable, non-invasive solution for ensuring the safety of customized LLMs. Our analysis across various fine-tuning scenarios confirms SafetyLock's robustness, advocating its integration into safety protocols for aligned LLMs. The code is released at https://github.com/zhu-minjun/SafetyLock.
Harmful fine-tuning attack poses a serious threat to the online fine-tuning service. Vaccine, a recent alignment-stage defense, applies uniform perturbation to all layers of embedding to make the model robust to the simulated embedding drift. However, applying layer-wise uniform perturbation may lead to excess perturbations for some particular safety-irrelevant layers, resulting in defense performance degradation and unnecessary memory consumption. To address this limitation, we propose Targeted Vaccine (T-Vaccine), a memory-efficient safety alignment method that applies perturbation to only selected layers of the model. T-Vaccine follows two core steps: First, it uses gradient norm as a statistical metric to identify the safety-critical layers. Second, instead of applying uniform perturbation across all layers, T-Vaccine only applies perturbation to the safety-critical layers while keeping other layers frozen during training. Results show that T-Vaccine outperforms Vaccine in terms of both defense effectiveness and resource efficiency. Comparison with other defense baselines, e.g., RepNoise and TAR also demonstrate the superiority of T-Vaccine. Notably, T-Vaccine is the first defense that can address harmful fine-tuning issues for a 7B pre-trained models trained on consumer GPUs with limited memory (e.g., RTX 4090). Our code is available at https://github.com/Lslland/T-Vaccine.
Large language models (LLMs) have gained human trust due to their capabilities and helpfulness. However, this in turn may allow LLMs to affect users' mindsets by manipulating language. It is termed as gaslighting, a psychological effect. In this work, we aim to investigate the vulnerability of LLMs under prompt-based and fine-tuning-based gaslighting attacks. Therefore, we propose a two-stage framework DeepCoG designed to: 1) elicit gaslighting plans from LLMs with the proposed DeepGaslighting prompting template, and 2) acquire gaslighting conversations from LLMs through our Chain-of-Gaslighting method. The gaslighting conversation dataset along with a corresponding safe dataset is applied to fine-tuning-based attacks on open-source LLMs and anti-gaslighting safety alignment on these LLMs. Experiments demonstrate that both prompt-based and fine-tuning-based attacks transform three open-source LLMs into gaslighters. In contrast, we advanced three safety alignment strategies to strengthen (by 12.05%) the safety guardrail of LLMs. Our safety alignment strategies have minimal impacts on the utility of LLMs. Empirical studies indicate that an LLM may be a potential gaslighter, even if it passed the harmfulness test on general dangerous queries.
This study exposes the safety vulnerabilities of Large Language Models (LLMs) in multi-turn interactions, where malicious users can obscure harmful intents across several queries. We introduce ActorAttack, a novel multi-turn attack method inspired by actor-network theory, which models a network of semantically linked actors as attack clues to generate diverse and effective attack paths toward harmful targets. ActorAttack addresses two main challenges in multi-turn attacks: (1) concealing harmful intents by creating an innocuous conversation topic about the actor, and (2) uncovering diverse attack paths towards the same harmful target by leveraging LLMs' knowledge to specify the correlated actors as various attack clues. In this way, ActorAttack outperforms existing single-turn and multi-turn attack methods across advanced aligned LLMs, even for GPT-o1. We will publish a dataset called SafeMTData, which includes multi-turn adversarial prompts and safety alignment data, generated by ActorAttack. We demonstrate that models safety-tuned using our safety dataset are more robust to multi-turn attacks. Code is available at https://github.com/renqibing/ActorAttack.
This paper studies the vulnerabilities of transformer-based Large Language Models (LLMs) to jailbreaking attacks, focusing specifically on the optimization-based Greedy Coordinate Gradient (GCG) strategy. We first observe a positive correlation between the effectiveness of attacks and the internal behaviors of the models. For instance, attacks tend to be less effective when models pay more attention to system prompts designed to ensure LLM safety alignment. Building on this discovery, we introduce an enhanced method that manipulates models' attention scores to facilitate LLM jailbreaking, which we term AttnGCG. Empirically, AttnGCG shows consistent improvements in attack efficacy across diverse LLMs, achieving an average increase of ~7% in the Llama-2 series and ~10% in the Gemma series. Our strategy also demonstrates robust attack transferability against both unseen harmful goals and black-box LLMs like GPT-3.5 and GPT-4. Moreover, we note our attention-score visualization is more interpretable, allowing us to gain better insights into how our targeted attention manipulation facilitates more effective jailbreaking. We release the code at https://github.com/UCSC-VLAA/AttnGCG-attack.
The safety alignment ability of Vision-Language Models (VLMs) is prone to be degraded by the integration of the vision module compared to its LLM backbone. We investigate this phenomenon, dubbed as ''safety alignment degradation'' in this paper, and show that the challenge arises from the representation gap that emerges when introducing vision modality to VLMs. In particular, we show that the representations of multi-modal inputs shift away from that of text-only inputs which represent the distribution that the LLM backbone is optimized for. At the same time, the safety alignment capabilities, initially developed within the textual embedding space, do not successfully transfer to this new multi-modal representation space. To reduce safety alignment degradation, we introduce Cross-Modality Representation Manipulation (CMRM), an inference time representation intervention method for recovering the safety alignment ability that is inherent in the LLM backbone of VLMs, while simultaneously preserving the functional capabilities of VLMs. The empirical results show that our framework significantly recovers the alignment ability that is inherited from the LLM backbone with minimal impact on the fluency and linguistic capabilities of pre-trained VLMs even without additional training. Specifically, the unsafe rate of LLaVA-7B on multi-modal input can be reduced from 61.53% to as low as 3.15% with only inference-time intervention. WARNING: This paper contains examples of toxic or harmful language.