To mitigate the risk of harmful outputs from large vision models (LVMs), we introduce the SafeSora dataset to promote research on aligning text-to-video generation with human values. This dataset encompasses human preferences in text-to-video generation tasks along two primary dimensions: helpfulness and harmlessness. To capture in-depth human preferences and facilitate structured reasoning by crowdworkers, we subdivide helpfulness into 4 sub-dimensions and harmlessness into 12 sub-categories, serving as the basis for pilot annotations. The SafeSora dataset includes 14,711 unique prompts, 57,333 unique videos generated by 4 distinct LVMs, and 51,691 pairs of preference annotations labeled by humans. We further demonstrate the utility of the SafeSora dataset through several applications, including training the text-video moderation model and aligning LVMs with human preference by fine-tuning a prompt augmentation module or the diffusion model. These applications highlight its potential as the foundation for text-to-video alignment research, such as human preference modeling and the development and validation of alignment algorithms.
Large language models (LLMs) excel in various capabilities but also pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment from the perspective of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose generation-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects. Experiments on multiple recent LLMs show that: (1) Safety neurons are sparse and effective. We can restore $90$% safety performance with intervention only on about $5$% of all the neurons. (2) Safety neurons encode transferrable mechanisms. They exhibit consistent effectiveness on different red-teaming datasets. The finding of safety neurons also interprets "alignment tax". We observe that the identified key neurons for safety and helpfulness significantly overlap, but they require different activation patterns of the shared neurons. Furthermore, we demonstrate an application of safety neurons in detecting unsafe outputs before generation. Our findings may promote further research on understanding LLM alignment. The source codes will be publicly released to facilitate future research.
Merging Large Language Models (LLMs) is a cost-effective technique for combining multiple expert LLMs into a single versatile model, retaining the expertise of the original ones. However, current approaches often overlook the importance of safety alignment during merging, leading to highly misaligned models. This work investigates the effects of model merging on alignment. We evaluate several popular model merging techniques, demonstrating that existing methods do not only transfer domain expertise but also propagate misalignment. We propose a simple two-step approach to address this problem: (i) generating synthetic safety and domain-specific data, and (ii) incorporating these generated data into the optimization process of existing data-aware model merging techniques. This allows us to treat alignment as a skill that can be maximized in the resulting merged LLM. Our experiments illustrate the effectiveness of integrating alignment-related data during merging, resulting in models that excel in both domain expertise and alignment.
Safety-aligned language models often exhibit fragile and imbalanced safety mechanisms, increasing the likelihood of generating unsafe content. In addition, incorporating new knowledge through editing techniques to language models can further compromise safety. To address these issues, we propose SafeInfer, a context-adaptive, decoding-time safety alignment strategy for generating safe responses to user queries. SafeInfer comprises two phases: the safety amplification phase, which employs safe demonstration examples to adjust the model's hidden states and increase the likelihood of safer outputs, and the safety-guided decoding phase, which influences token selection based on safety-optimized distributions, ensuring the generated content complies with ethical guidelines. Further, we present HarmEval, a novel benchmark for extensive safety evaluations, designed to address potential misuse scenarios in accordance with the policies of leading AI tech giants.
The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness.
Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering. Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content. We propose Safety Arithmetic, a training-free framework enhancing LLM safety across different scenarios: Base models, Supervised fine-tuned models (SFT), and Edited models. Safety Arithmetic involves Harm Direction Removal to avoid harmful content and Safety Alignment to promote safe responses. Additionally, we present NoIntentEdit, a dataset highlighting edit instances that could compromise model safety if used unintentionally. Our experiments show that Safety Arithmetic significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation.
Federated learning (FL) enables multiple parties to collaboratively fine-tune an large language model (LLM) without the need of direct data sharing. Ideally, by training on decentralized data that is aligned with human preferences and safety principles, federated instruction tuning can result in an LLM that could behave in a helpful and safe manner. In this paper, we for the first time reveal the vulnerability of safety alignment in FedIT by proposing a simple, stealthy, yet effective safety attack method. Specifically, the malicious clients could automatically generate attack data without involving manual efforts and attack the FedIT system by training their local LLMs on such attack data. Unfortunately, this proposed safety attack not only can compromise the safety alignment of LLM trained via FedIT, but also can not be effectively defended against by many existing FL defense methods. Targeting this, we further propose a post-hoc defense method, which could rely on a fully automated pipeline: generation of defense data and further fine-tuning of the LLM. Extensive experiments show that our safety attack method can significantly compromise the LLM's safety alignment (e.g., reduce safety rate by 70\%), which can not be effectively defended by existing defense methods (at most 4\% absolute improvement), while our safety defense method can significantly enhance the attacked LLM's safety alignment (at most 69\% absolute improvement).
Large language models (LLMs) rely on safety alignment to avoid responding to malicious user inputs. Unfortunately, jailbreak can circumvent safety guardrails, resulting in LLMs generating harmful content and raising concerns about LLM safety. Due to language models with intensive parameters often regarded as black boxes, the mechanisms of alignment and jailbreak are challenging to elucidate. In this paper, we employ weak classifiers to explain LLM safety through the intermediate hidden states. We first confirm that LLMs learn ethical concepts during pre-training rather than alignment and can identify malicious and normal inputs in the early layers. Alignment actually associates the early concepts with emotion guesses in the middle layers and then refines them to the specific reject tokens for safe generations. Jailbreak disturbs the transformation of early unethical classification into negative emotions. We conduct experiments on models from 7B to 70B across various model families to prove our conclusion. Overall, our paper indicates the intrinsical mechanism of LLM safety and how jailbreaks circumvent safety guardrails, offering a new perspective on LLM safety and reducing concerns. Our code is available at https://github.com/ydyjya/LLM-IHS-Explanation.
Jailbreaking is an emerging adversarial attack that bypasses the safety alignment deployed in off-the-shelf large language models (LLMs) and has evolved into multiple categories: human-based, optimization-based, generation-based, and the recent indirect and multilingual jailbreaks. However, delivering a practical jailbreak defense is challenging because it needs to not only handle all the above jailbreak attacks but also incur negligible delays to user prompts, as well as be compatible with both open-source and closed-source LLMs. Inspired by how the traditional security concept of shadow stacks defends against memory overflow attacks, this paper introduces a generic LLM jailbreak defense framework called SelfDefend, which establishes a shadow LLM as a defense instance (in detection state) to concurrently protect the target LLM instance (in normal answering state) in the normal stack and collaborate with it for checkpoint-based access control. The effectiveness of SelfDefend builds upon our observation that existing LLMs can identify harmful prompts or intentions in user queries, which we empirically validate using mainstream GPT-3.5/4 models against major jailbreak attacks. To further improve the defense's robustness and minimize costs, we employ a data distillation approach to tune dedicated open-source defense models. When deployed to protect GPT-3.5/4, Claude, Llama-2-7b/13b, and Mistral, these models outperform seven state-of-the-art defenses and match the performance of GPT-4-based SelfDefend, with significantly lower extra delays. Further experiments show that the tuned models are robust to adaptive jailbreaks and prompt injections.
The safety alignment of current Large Language Models (LLMs) is vulnerable. Relatively simple attacks, or even benign fine-tuning, can jailbreak aligned models. We argue that many of these vulnerabilities are related to a shared underlying issue: safety alignment can take shortcuts, wherein the alignment adapts a model's generative distribution primarily over only its very first few output tokens. We refer to this issue as shallow safety alignment. In this paper, we present case studies to explain why shallow safety alignment can exist and provide evidence that current aligned LLMs are subject to this issue. We also show how these findings help explain multiple recently discovered vulnerabilities in LLMs, including the susceptibility to adversarial suffix attacks, prefilling attacks, decoding parameter attacks, and fine-tuning attacks. Importantly, we discuss how this consolidated notion of shallow safety alignment sheds light on promising research directions for mitigating these vulnerabilities. For instance, we show that deepening the safety alignment beyond just the first few tokens can often meaningfully improve robustness against some common exploits. Finally, we design a regularized finetuning objective that makes the safety alignment more persistent against fine-tuning attacks by constraining updates on initial tokens. Overall, we advocate that future safety alignment should be made more than just a few tokens deep.
Recent studies reveal that integrating new modalities into Large Language Models (LLMs), such as Vision-Language Models (VLMs), creates a new attack surface that bypasses existing safety training techniques like Supervised Fine-tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF). While further SFT and RLHF-based safety training can be conducted in multi-modal settings, collecting multi-modal training datasets poses a significant challenge. Inspired by the structural design of recent multi-modal models, where, regardless of the combination of input modalities, all inputs are ultimately fused into the language space, we aim to explore whether unlearning solely in the textual domain can be effective for cross-modality safety alignment. Our evaluation across six datasets empirically demonstrates the transferability -- textual unlearning in VLMs significantly reduces the Attack Success Rate (ASR) to less than 8\% and in some cases, even as low as nearly 2\% for both text-based and vision-text-based attacks, alongside preserving the utility. Moreover, our experiments show that unlearning with a multi-modal dataset offers no potential benefits but incurs significantly increased computational demands, possibly up to 6 times higher.
Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only the task's goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. In this paper, we unveil that intrinsic self-correction can be progressively improved, allowing it to approach a converged state. Our findings are verified in: (1) the scenario of multi-round question answering, by comprehensively demonstrating that intrinsic self-correction can progressively introduce performance gains through iterative interactions, ultimately converging to stable performance; and (2) the context of intrinsic self-correction for enhanced morality, in which we provide empirical evidence that iteratively applying instructions reduces model uncertainty towards convergence, which then leads to convergence of both the calibration error and self-correction performance, ultimately resulting in a stable state of intrinsic self-correction. Furthermore, we introduce a mathematical formulation and a simulation task indicating that the latent concepts activated by self-correction instructions drive the reduction of model uncertainty. Based on our experimental results and analysis of the convergence of intrinsic self-correction, we reveal its underlying mechanism: consistent injected instructions reduce model uncertainty which yields converged, improved performance.
Large Language Models (LLMs) require careful safety alignment to prevent malicious outputs. While significant research focuses on mitigating harmful content generation, the enhanced safety often come with the side effect of over-refusal, where LLMs may reject innocuous prompts and become less helpful. Although the issue of over-refusal has been empirically observed, a systematic measurement is challenging due to the difficulty of crafting prompts that can elicit the over-refusal behaviors of LLMs. This study proposes a novel method for automatically generating large-scale over-refusal datasets. Leveraging this technique, we introduce OR-Bench, the first large-scale over-refusal benchmark. OR-Bench comprises 80,000 over-refusal prompts across 10 common rejection categories, a subset of around 1,000 hard prompts that are challenging even for state-of-the-art LLMs, and an additional 600 toxic prompts to prevent indiscriminate responses. We then conduct a comprehensive study to measure the over-refusal of 32 popular LLMs across 8 model families. Our datasets are publicly available at https://huggingface.co/bench-llms and our codebase is open-sourced at https://github.com/justincui03/or-bench. We hope this benchmark can help the community develop better safety aligned models.
Along with the remarkable successes of Language language models, recent research also started to explore the security threats of LLMs, including jailbreaking attacks. Attackers carefully craft jailbreaking prompts such that a target LLM will respond to the harmful question. Existing jailbreaking attacks require either human experts or leveraging complicated algorithms to craft jailbreaking prompts. In this paper, we introduce BOOST, a simple attack that leverages only the eos tokens. We demonstrate that rather than constructing complicated jailbreaking prompts, the attacker can simply append a few eos tokens to the end of a harmful question. It will bypass the safety alignment of LLMs and lead to successful jailbreaking attacks. We further apply BOOST to four representative jailbreak methods and show that the attack success rates of these methods can be significantly enhanced by simply adding eos tokens to the prompt. To understand this simple but novel phenomenon, we conduct empirical analyses. Our analysis reveals that adding eos tokens makes the target LLM believe the input is much less harmful, and eos tokens have low attention values and do not affect LLM's understanding of the harmful questions, leading the model to actually respond to the questions. Our findings uncover how fragile an LLM is against jailbreak attacks, motivating the development of strong safety alignment approaches.
The growing safety concerns surrounding large language models raise an urgent need to align them with diverse human preferences to simultaneously enhance their helpfulness and safety. A promising approach is to enforce safety constraints through Reinforcement Learning from Human Feedback (RLHF). For such constrained RLHF, typical Lagrangian-based primal-dual policy optimization methods are computationally expensive and often unstable. This paper presents a perspective of dualization that reduces constrained alignment to an equivalent unconstrained alignment problem. We do so by pre-optimizing a smooth and convex dual function that has a closed form. This shortcut eliminates the need for cumbersome primal-dual policy iterations, greatly reducing the computational burden and improving training stability. Our strategy leads to two practical algorithms in model-based and preference-based settings (MoCAN and PeCAN, respectively). A broad range of experiments demonstrate the effectiveness and merits of our algorithms.
Large language models (LLMs) are vulnerable when trained on datasets containing harmful content, which leads to potential jailbreaking attacks in two scenarios: the integration of harmful texts within crowdsourced data used for pre-training and direct tampering with LLMs through fine-tuning. In both scenarios, adversaries can compromise the safety alignment of LLMs, exacerbating malfunctions. Motivated by the need to mitigate these adversarial influences, our research aims to enhance safety alignment by either neutralizing the impact of malicious texts in pre-training datasets or increasing the difficulty of jailbreaking during downstream fine-tuning. In this paper, we propose a data curation framework designed to counter adversarial impacts in both scenarios. Our method operates under the assumption that we have no prior knowledge of attack details, focusing solely on curating clean texts. We introduce an iterative process aimed at revising texts to reduce their perplexity as perceived by LLMs, while simultaneously preserving their text quality. By pre-training or fine-tuning LLMs with curated clean texts, we observe a notable improvement in LLM robustness regarding safety alignment against harmful queries. For instance, when pre-training LLMs using a crowdsourced dataset containing 5\% harmful instances, adding an equivalent amount of curated texts significantly mitigates the likelihood of providing harmful responses in LLMs and reduces the attack success rate by 71\%. Our study represents a significant step towards mitigating the risks associated with training-based jailbreaking and fortifying the secure utilization of LLMs.
Recent studies show that Large Language Models (LLMs) with safety alignment can be jail-broken by fine-tuning on a dataset mixed with harmful data. First time in the literature, we show that the jail-broken effect can be mitigated by separating states in the finetuning stage to optimize the alignment and user datasets. Unfortunately, our subsequent study shows that this simple Bi-State Optimization (BSO) solution experiences convergence instability when steps invested in its alignment state is too small, leading to downgraded alignment performance. By statistical analysis, we show that the \textit{excess drift} towards consensus could be a probable reason for the instability. To remedy this issue, we propose \textbf{L}azy(\textbf{i}) \textbf{s}afety \textbf{a}lignment (\textbf{Lisa}), which introduces a proximal term to constraint the drift of each state. Theoretically, the benefit of the proximal term is supported by the convergence analysis, wherein we show that a sufficient large proximal factor is necessary to guarantee Lisa's convergence. Empirically, our results on four downstream finetuning tasks show that Lisa with a proximal term can significantly increase alignment performance while maintaining the LLM's accuracy on the user tasks. Code is available at \url{https://github.com/git-disl/Lisa}.
The existing safety alignment of Large Language Models (LLMs) is found fragile and could be easily attacked through different strategies, such as through fine-tuning on a few harmful examples or manipulating the prefix of the generation results. However, the attack mechanisms of these strategies are still underexplored. In this paper, we ask the following question: \textit{while these approaches can all significantly compromise safety, do their attack mechanisms exhibit strong similarities?} To answer this question, we break down the safeguarding process of an LLM when encountered with harmful instructions into three stages: (1) recognizing harmful instructions, (2) generating an initial refusing tone, and (3) completing the refusal response. Accordingly, we investigate whether and how different attack strategies could influence each stage of this safeguarding process. We utilize techniques such as logit lens and activation patching to identify model components that drive specific behavior, and we apply cross-model probing to examine representation shifts after an attack. In particular, we analyze the two most representative types of attack approaches: Explicit Harmful Attack (EHA) and Identity-Shifting Attack (ISA). Surprisingly, we find that their attack mechanisms diverge dramatically. Unlike ISA, EHA tends to aggressively target the harmful recognition stage. While both EHA and ISA disrupt the latter two stages, the extent and mechanisms of their attacks differ significantly. Our findings underscore the importance of understanding LLMs' internal safeguarding process and suggest that diverse defense mechanisms are required to effectively cope with various types of attacks.
Safety alignment is crucial to ensure that large language models (LLMs) behave in ways that align with human preferences and prevent harmful actions during inference. However, recent studies show that the alignment can be easily compromised through finetuning with only a few adversarially designed training examples. We aim to measure the risks in finetuning LLMs through navigating the LLM safety landscape. We discover a new phenomenon observed universally in the model parameter space of popular open-source LLMs, termed as "safety basin": random perturbations to model weights maintain the safety level of the original aligned model within its local neighborhood. However, outside this local region, safety is fully compromised, exhibiting a sharp, step-like drop. This safety basin contrasts sharply with the LLM capability landscape, where model performance peaks at the origin and gradually declines as random perturbation increases. Our discovery inspires us to propose the new VISAGE safety metric that measures the safety in LLM finetuning by probing its safety landscape. Visualizing the safety landscape of the aligned model enables us to understand how finetuning compromises safety by dragging the model away from the safety basin. The LLM safety landscape also highlights the system prompt's critical role in protecting a model, and that such protection transfers to its perturbed variants within the safety basin. These observations from our safety landscape research provide new insights for future work on LLM safety community. Our code is publicly available at https://github.com/ShengYun-Peng/llm-landscape.
Safety alignment of large language models (LLMs) has been gaining increasing attention. However, current safety-aligned LLMs suffer from the fragile and imbalanced safety mechanisms, which can still be induced to generate unsafe responses, exhibit over-safety by rejecting safe user inputs, and fail to preserve general utility after safety alignment. To this end, we propose a novel post safety alignment (PSA) method to address these inherent and emerging safety challenges, including safety enhancement, over-safety mitigation, and utility preservation. In specific, we introduce \textsc{SafePatching}, a novel framework for comprehensive PSA, where two distinct safety patches are developed on the harmful data to enhance safety and mitigate over-safety concerns, and then seamlessly integrated into the target LLM backbone without compromising its utility. Extensive experiments on four representative aligned LLMs, including LLaMA-2/3, Gemma and Mistral, show that \textsc{SafePatching} achieves a more comprehensive PSA than baseline methods, further optimizing the balance between being helpful and harmless in current aligned LLMs. Also, \textsc{SafePatching} demonstrates its superiority in continual PSA scenarios.