Conventional language model (LM) safety alignment relies on a reactive, disjoint procedure: attackers exploit a static model, followed by defensive fine-tuning to patch exposed vulnerabilities. This sequential approach creates a mismatch -- attackers overfit to obsolete defenses, while defenders perpetually lag behind emerging threats. To address this, we propose Self-RedTeam, an online self-play reinforcement learning algorithm where an attacker and defender agent co-evolve through continuous interaction. We cast safety alignment as a two-player zero-sum game, where a single model alternates between attacker and defender roles -- generating adversarial prompts and safeguarding against them -- while a reward LM adjudicates outcomes. This enables dynamic co-adaptation. Grounded in the game-theoretic framework of zero-sum games, we establish a theoretical safety guarantee which motivates the design of our method: if self-play converges to a Nash Equilibrium, the defender will reliably produce safe responses to any adversarial input. Empirically, Self-RedTeam uncovers more diverse attacks (+21.8% SBERT) compared to attackers trained against static defenders and achieves higher robustness on safety benchmarks (e.g., +65.5% on WildJailBreak) than defenders trained against static attackers. We further propose hidden Chain-of-Thought, allowing agents to plan privately, which boosts adversarial diversity and reduces over-refusals. Our results motivate a shift from reactive patching to proactive co-evolution in LM safety training, enabling scalable, autonomous, and robust self-improvement of LMs via multi-agent reinforcement learning (MARL).
Adversarial threats against LLMs are escalating faster than current defenses can adapt. We expose a critical geometric blind spot in alignment: adversarial prompts exploit latent camouflage, embedding perilously close to the safe representation manifold while encoding unsafe intent thereby evading surface level defenses like Direct Preference Optimization (DPO), which remain blind to the latent geometry. We introduce ALKALI, the first rigorously curated adversarial benchmark and the most comprehensive to date spanning 9,000 prompts across three macro categories, six subtypes, and fifteen attack families. Evaluation of 21 leading LLMs reveals alarmingly high Attack Success Rates (ASRs) across both open and closed source models, exposing an underlying vulnerability we term latent camouflage, a structural blind spot where adversarial completions mimic the latent geometry of safe ones. To mitigate this vulnerability, we introduce GRACE - Geometric Representation Aware Contrastive Enhancement, an alignment framework coupling preference learning with latent space regularization. GRACE enforces two constraints: latent separation between safe and adversarial completions, and adversarial cohesion among unsafe and jailbreak behaviors. These operate over layerwise pooled embeddings guided by a learned attention profile, reshaping internal geometry without modifying the base model, and achieve up to 39% ASR reduction. Moreover, we introduce AVQI, a geometry aware metric that quantifies latent alignment failure via cluster separation and compactness. AVQI reveals when unsafe completions mimic the geometry of safe ones, offering a principled lens into how models internally encode safety. We make the code publicly available at https://anonymous.4open.science/r/alkali-B416/README.md.
Harmful fine-tuning (HFT), performed directly on open-source LLMs or through Fine-tuning-as-a-Service, breaks safety alignment and poses significant threats. Existing methods aim to mitigate HFT risks by learning robust representation on alignment data or making harmful data unlearnable, but they treat each data sample equally, leaving data vulnerability patterns understudied. In this work, we reveal that certain subsets of alignment data are consistently more prone to forgetting during HFT across different fine-tuning tasks. Inspired by these findings, we propose Vulnerability-Aware Alignment (VAA), which estimates data vulnerability, partitions data into "vulnerable" and "invulnerable" groups, and encourages balanced learning using a group distributionally robust optimization (Group DRO) framework. Specifically, VAA learns an adversarial sampler that samples examples from the currently underperforming group and then applies group-dependent adversarial perturbations to the data during training, aiming to encourage a balanced learning process across groups. Experiments across four fine-tuning tasks demonstrate that VAA significantly reduces harmful scores while preserving downstream task performance, outperforming state-of-the-art baselines.
To construct responsible and secure AI applications, harmful information data is widely utilized for adversarial testing and the development of safeguards. Existing studies mainly leverage Large Language Models (LLMs) to synthesize data to obtain high-quality task datasets at scale, thereby avoiding costly human annotation. However, limited by the safety alignment mechanisms of LLMs, the synthesis of harmful data still faces challenges in generation reliability and content diversity. In this study, we propose a novel harmful information synthesis framework, PoisonSwarm, which applies the model crowdsourcing strategy to generate diverse harmful data while maintaining a high success rate. Specifically, we generate abundant benign data as the based templates in a counterfactual manner. Subsequently, we decompose each based template into multiple semantic units and perform unit-by-unit toxification and final refinement through dynamic model switching, thus ensuring the success of synthesis. Experimental results demonstrate that PoisonSwarm achieves state-of-the-art performance in synthesizing different categories of harmful data with high scalability and diversity.
The inherent risk of generating harmful and unsafe content by Large Language Models (LLMs), has highlighted the need for their safety alignment. Various techniques like supervised fine-tuning, reinforcement learning from human feedback, and red-teaming were developed for ensuring the safety alignment of LLMs. However, the robustness of these aligned LLMs is always challenged by adversarial attacks that exploit unexplored and underlying vulnerabilities of the safety alignment. In this paper, we develop a novel black-box jailbreak attack, called BitBypass, that leverages hyphen-separated bitstream camouflage for jailbreaking aligned LLMs. This represents a new direction in jailbreaking by exploiting fundamental information representation of data as continuous bits, rather than leveraging prompt engineering or adversarial manipulations. Our evaluation of five state-of-the-art LLMs, namely GPT-4o, Gemini 1.5, Claude 3.5, Llama 3.1, and Mixtral, in adversarial perspective, revealed the capabilities of BitBypass in bypassing their safety alignment and tricking them into generating harmful and unsafe content. Further, we observed that BitBypass outperforms several state-of-the-art jailbreak attacks in terms of stealthiness and attack success. Overall, these results highlights the effectiveness and efficiency of BitBypass in jailbreaking these state-of-the-art LLMs.
Large Language Models (LLMs) have evolved into Multimodal Large Language Models (MLLMs), significantly enhancing their capabilities by integrating visual information and other types, thus aligning more closely with the nature of human intelligence, which processes a variety of data forms beyond just text. Despite advancements, the undesirable generation of these models remains a critical concern, particularly due to vulnerabilities exposed by text-based jailbreak attacks, which have represented a significant threat by challenging existing safety protocols. Motivated by the unique security risks posed by the integration of new and old modalities for MLLMs, we propose a unified multimodal universal jailbreak attack framework that leverages iterative image-text interactions and transfer-based strategy to generate a universal adversarial suffix and image. Our work not only highlights the interaction of image-text modalities can be used as a critical vulnerability but also validates that multimodal universal jailbreak attacks can bring higher-quality undesirable generations across different MLLMs. We evaluate the undesirable context generation of MLLMs like LLaVA, Yi-VL, MiniGPT4, MiniGPT-v2, and InstructBLIP, and reveal significant multimodal safety alignment issues, highlighting the inadequacy of current safety mechanisms against sophisticated multimodal attacks. This study underscores the urgent need for robust safety measures in MLLMs, advocating for a comprehensive review and enhancement of security protocols to mitigate potential risks associated with multimodal capabilities.
Fine-tuning large language models (LLMs) for telecom tasks and datasets is a common practice to adapt general-purpose models to the telecom domain. However, little attention has been paid to how this process may compromise model safety. Recent research has shown that even benign fine-tuning can degrade the safety alignment of LLMs, causing them to respond to harmful or unethical user queries. In this paper, we investigate this issue for telecom-tuned LLMs using three representative datasets featured by the GenAINet initiative. We show that safety degradation persists even for structured and seemingly harmless datasets such as 3GPP standards and tabular records, indicating that telecom-specific data is not immune to safety erosion during fine-tuning. We further extend our analysis to publicly available Telecom LLMs trained via continual pre-training, revealing that safety alignment is often severely lacking, primarily due to the omission of safety-focused instruction tuning. To address these issues in both fine-tuned and pre-trained models, we conduct extensive experiments and evaluate three safety realignment defenses (SafeInstruct, SafeLoRA, and SafeMERGE) using established red-teaming benchmarks. The results show that, across all settings, the proposed defenses can effectively restore safety after harmful degradation without compromising downstream task performance, leading to Safe teleCOMMunication (SafeCOMM) models. In a nutshell, our work serves as a diagnostic study and practical guide for safety realignment in telecom-tuned LLMs, and emphasizes the importance of safety-aware instruction and fine-tuning for real-world deployments of Telecom LLMs.
Recent advancements in large language models (LLMs) have underscored their vulnerability to safety alignment jailbreaks, particularly when subjected to downstream fine-tuning. However, existing mitigation strategies primarily focus on reactively addressing jailbreak incidents after safety guardrails have been compromised, removing harmful gradients during fine-tuning, or continuously reinforcing safety alignment throughout fine-tuning. As such, they tend to overlook a critical upstream factor: the role of the original safety-alignment data. This paper therefore investigates the degradation of safety guardrails through the lens of representation similarity between upstream alignment datasets and downstream fine-tuning tasks. Our experiments demonstrate that high similarity between these datasets significantly weakens safety guardrails, making models more susceptible to jailbreaks. Conversely, low similarity between these two types of datasets yields substantially more robust models and thus reduces harmfulness score by up to 10.33%. By highlighting the importance of upstream dataset design in the building of durable safety guardrails and reducing real-world vulnerability to jailbreak attacks, these findings offer actionable insights for fine-tuning service providers.
Large Language Models (LLMs) excel in various natural language processing tasks but remain vulnerable to generating harmful content or being exploited for malicious purposes. Although safety alignment datasets have been introduced to mitigate such risks through supervised fine-tuning (SFT), these datasets often lack comprehensive risk coverage. Most existing datasets focus primarily on lexical diversity while neglecting other critical dimensions. To address this limitation, we propose a novel analysis framework to systematically measure the risk coverage of alignment datasets across three essential dimensions: Lexical Diversity, Malicious Intent, and Jailbreak Tactics. We further introduce TRIDENT, an automated pipeline that leverages persona-based, zero-shot LLM generation to produce diverse and comprehensive instructions spanning these dimensions. Each harmful instruction is paired with an ethically aligned response, resulting in two datasets: TRIDENT-Core, comprising 26,311 examples, and TRIDENT-Edge, with 18,773 examples. Fine-tuning Llama 3.1-8B on TRIDENT-Edge demonstrates substantial improvements, achieving an average 14.29% reduction in Harm Score, and a 20% decrease in Attack Success Rate compared to the best-performing baseline model fine-tuned on the WildBreak dataset.
Ensuring Vision-Language Models (VLMs) generate safe outputs is crucial for their reliable deployment. However, LVLMs suffer from drastic safety degradation compared to their LLM backbone. Even blank or irrelevant images can trigger LVLMs to generate harmful responses to prompts that would otherwise be refused in text-only contexts. The modality gap between image and text representations has been recently hypothesized to contribute to safety degradation of LVLMs. However, if and how the amount of modality gap affects LVLMs' safety is not studied. In this work, we show that the amount of modality gap is highly inversely correlated with VLMs' safety. Then, we show that this modality gap is introduced during pretraining LVLMs and persists through fine-tuning. Inspired by this observation, we propose a regularization to reduce the modality gap during pretraining. Our extensive experiments on LLaVA v1.5, ShareGPT4V, and MiniGPT-4 show that our method substantially improves safety alignment of LVLMs, reducing unsafe rate by up to 16.3% without compromising performance, and can further boost existing defenses by up to 18.2%.
Large Vision-Language Models (LVLMs) have shown remarkable capabilities across a wide range of multimodal tasks. However, their integration of visual inputs introduces expanded attack surfaces, thereby exposing them to novel security vulnerabilities. In this work, we conduct a systematic representational analysis to uncover why conventional adversarial attacks can circumvent the safety mechanisms embedded in LVLMs. We further propose a novel two stage evaluation framework for adversarial attacks on LVLMs. The first stage differentiates among instruction non compliance, outright refusal, and successful adversarial exploitation. The second stage quantifies the degree to which the model's output fulfills the harmful intent of the adversarial prompt, while categorizing refusal behavior into direct refusals, soft refusals, and partial refusals that remain inadvertently helpful. Finally, we introduce a normative schema that defines idealized model behavior when confronted with harmful prompts, offering a principled target for safety alignment in multimodal systems.
Large language models (LLMs) frequently refuse to respond to pseudo-malicious instructions: semantically harmless input queries triggering unnecessary LLM refusals due to conservative safety alignment, significantly impairing user experience. Collecting such instructions is crucial for evaluating and mitigating over-refusals, but existing instruction curation methods, like manual creation or instruction rewriting, either lack scalability or fail to produce sufficiently diverse and effective refusal-inducing prompts. To address these limitations, we introduce EVOREFUSE, a prompt optimization approach that generates diverse pseudo-malicious instructions consistently eliciting confident refusals across LLMs. EVOREFUSE employs an evolutionary algorithm exploring the instruction space in more diverse directions than existing methods via mutation strategies and recombination, and iteratively evolves seed instructions to maximize evidence lower bound on LLM refusal probability. Using EVOREFUSE, we create two novel datasets: EVOREFUSE-TEST, a benchmark of 582 pseudo-malicious instructions that outperforms the next-best benchmark with 140.41% higher average refusal triggering rate across 9 LLMs, 34.86% greater lexical diversity, and 40.03% improved LLM response confidence scores; and EVOREFUSE-ALIGN, which provides 3,000 pseudo-malicious instructions with responses for supervised and preference-based alignment training. LLAMA3.1-8B-INSTRUCT supervisedly fine-tuned on EVOREFUSE-ALIGN achieves up to 14.31% fewer over-refusals than models trained on the second-best alignment dataset, without compromising safety. Our analysis with EVOREFUSE-TEST reveals models trigger over-refusals by overly focusing on sensitive keywords while ignoring broader context.
The acquisition of agentic capabilities has transformed LLMs from "knowledge providers" to "action executors", a trend that while expanding LLMs' capability boundaries, significantly increases their susceptibility to malicious use. Previous work has shown that current LLM-based agents execute numerous malicious tasks even without being attacked, indicating a deficiency in agentic use safety alignment during the post-training phase. To address this gap, we propose AgentAlign, a novel framework that leverages abstract behavior chains as a medium for safety alignment data synthesis. By instantiating these behavior chains in simulated environments with diverse tool instances, our framework enables the generation of highly authentic and executable instructions while capturing complex multi-step dynamics. The framework further ensures model utility by proportionally synthesizing benign instructions through non-malicious interpretations of behavior chains, precisely calibrating the boundary between helpfulness and harmlessness. Evaluation results on AgentHarm demonstrate that fine-tuning three families of open-source models using our method substantially improves their safety (35.8% to 79.5% improvement) while minimally impacting or even positively enhancing their helpfulness, outperforming various prompting methods. The dataset and code have both been open-sourced.
Large language models (LLMs) exhibit advancing capabilities in complex tasks, such as reasoning and graduate-level question answering, yet their resilience against misuse, particularly involving scientifically sophisticated risks, remains underexplored. Existing safety benchmarks typically focus either on instructions requiring minimal knowledge comprehension (e.g., ``tell me how to build a bomb") or utilize prompts that are relatively low-risk (e.g., multiple-choice or classification tasks about hazardous content). Consequently, they fail to adequately assess model safety when handling knowledge-intensive, hazardous scenarios. To address this critical gap, we introduce SOSBench, a regulation-grounded, hazard-focused benchmark encompassing six high-risk scientific domains: chemistry, biology, medicine, pharmacology, physics, and psychology. The benchmark comprises 3,000 prompts derived from real-world regulations and laws, systematically expanded via an LLM-assisted evolutionary pipeline that introduces diverse, realistic misuse scenarios (e.g., detailed explosive synthesis instructions involving advanced chemical formulas). We evaluate frontier models within a unified evaluation framework using our SOSBench. Despite their alignment claims, advanced models consistently disclose policy-violating content across all domains, demonstrating alarmingly high rates of harmful responses (e.g., 79.1% for Deepseek-R1 and 47.3% for GPT-4.1). These results highlight significant safety alignment deficiencies and underscore urgent concerns regarding the responsible deployment of powerful LLMs.
Text-to-Image (T2I) models have achieved remarkable success in generating visual content from text inputs. Although multiple safety alignment strategies have been proposed to prevent harmful outputs, they often lead to overly cautious behavior -- rejecting even benign prompts -- a phenomenon known as $\textit{over-refusal}$ that reduces the practical utility of T2I models. Despite over-refusal having been observed in practice, there is no large-scale benchmark that systematically evaluates this phenomenon for T2I models. In this paper, we present an automatic workflow to construct synthetic evaluation data, resulting in OVERT ($\textbf{OVE}$r-$\textbf{R}$efusal evaluation on $\textbf{T}$ext-to-image models), the first large-scale benchmark for assessing over-refusal behaviors in T2I models. OVERT includes 4,600 seemingly harmful but benign prompts across nine safety-related categories, along with 1,785 genuinely harmful prompts (OVERT-unsafe) to evaluate the safety-utility trade-off. Using OVERT, we evaluate several leading T2I models and find that over-refusal is a widespread issue across various categories (Figure 1), underscoring the need for further research to enhance the safety alignment of T2I models without compromising their functionality.As a preliminary attempt to reduce over-refusal, we explore prompt rewriting; however, we find it often compromises faithfulness to the meaning of the original prompts. Finally, we demonstrate the flexibility of our generation framework in accommodating diverse safety requirements by generating customized evaluation data adapting to user-defined policies.
Diffusion models have emerged as leading generative models for images and other modalities, but aligning their outputs with human preferences and safety constraints remains a critical challenge. This thesis proposal investigates methods to align diffusion models using reinforcement learning (RL) and reward modeling. We survey recent advances in fine-tuning text-to-image diffusion models with human feedback, including reinforcement learning from human and AI feedback, direct preference optimization, and differentiable reward approaches. We classify these methods based on the type of feedback (human, automated, binary or ranked preferences), the fine-tuning technique (policy gradient, reward-weighted likelihood, direct backpropagation, etc.), and their efficiency and safety outcomes. We compare key algorithms and frameworks, highlighting how they improve alignment with user intent or safety standards, and discuss inter-relationships such as how newer methods build on or diverge from earlier ones. Based on the survey, we identify five promising research directions for the next two years: (1) multi-objective alignment with combined rewards, (2) efficient human feedback usage and active learning, (3) robust safety alignment against adversarial inputs, (4) continual and online alignment of diffusion models, and (5) interpretable and trustworthy reward modeling for generative images. Each direction is elaborated with its problem statement, challenges, related work, and a proposed research plan. The proposal is organized as a comprehensive document with literature review, comparative tables of methods, and detailed research plans, aiming to contribute new insights and techniques for safer and value-aligned diffusion-based generative AI.
Finetuning large language models (LLMs) enables user-specific customization but introduces critical safety risks: even a few harmful examples can compromise safety alignment. A common mitigation strategy is to update the model more strongly on examples deemed safe, while downweighting or excluding those flagged as unsafe. However, because safety context can shift within a single example, updating the model equally on both harmful and harmless parts of a response is suboptimal-a coarse treatment we term static safety shaping. In contrast, we propose dynamic safety shaping (DSS), a framework that uses fine-grained safety signals to reinforce learning from safe segments of a response while suppressing unsafe content. To enable such fine-grained control during finetuning, we introduce a key insight: guardrail models, traditionally used for filtering, can be repurposed to evaluate partial responses, tracking how safety risk evolves throughout the response, segment by segment. This leads to the Safety Trajectory Assessment of Response (STAR), a token-level signal that enables shaping to operate dynamically over the training sequence. Building on this, we present STAR-DSS, guided by STAR scores, that robustly mitigates finetuning risks and delivers substantial safety improvements across diverse threats, datasets, and model families-all without compromising capability on intended tasks. We encourage future safety research to build on dynamic shaping principles for stronger mitigation against evolving finetuning risks.
Despite the remarkable proficiency of \textit{Large Reasoning Models} (LRMs) in handling complex reasoning tasks, their reliability in safety-critical scenarios remains uncertain. Existing evaluations primarily assess response-level safety, neglecting a critical issue we identify as \textbf{\textit{Superficial Safety Alignment} (SSA)} -- a phenomenon where models produce superficially safe outputs while internal reasoning processes fail to genuinely detect and mitigate underlying risks, resulting in inconsistent safety behaviors across multiple sampling attempts. To systematically investigate SSA, we introduce \textbf{Beyond Safe Answers (BSA)} bench, a novel benchmark comprising 2,000 challenging instances organized into three distinct SSA scenario types and spanning nine risk categories, each meticulously annotated with risk rationales. Evaluations of 19 state-of-the-art LRMs demonstrate the difficulty of this benchmark, with top-performing models achieving only 38.0\% accuracy in correctly identifying risk rationales. We further explore the efficacy of safety rules, specialized fine-tuning on safety reasoning data, and diverse decoding strategies in mitigating SSA. Our work provides a comprehensive assessment tool for evaluating and improving safety reasoning fidelity in LRMs, advancing the development of genuinely risk-aware and reliably safe AI systems.
Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they often refuse to answer legitimate queries-a phenomenon known as overrefusal. Overrefusal typically stems from over-conservative safety alignment, causing models to treat many reasonable prompts as potentially risky. To systematically understand this issue, we probe and leverage the models'safety decision boundaries to analyze and mitigate overrefusal. Our findings reveal that overrefusal is closely tied to misalignment at these boundary regions, where models struggle to distinguish subtle differences between benign and harmful content. Building on these insights, we present RASS, an automated framework for prompt generation and selection that strategically targets overrefusal prompts near the safety boundary. By harnessing steering vectors in the representation space, RASS efficiently identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. This approach not only provides a more precise and interpretable view of model safety decisions but also seamlessly extends to multilingual scenarios.We have explored the safety decision boundaries of various LLMs and construct the MORBench evaluation set to facilitate robust assessment of model safety and helpfulness across multiple languages. Code and datasets will be released at https://anonymous.4open.science/r/RASS-80D3.
As Large Language Models (LLMs) continue to advance and find applications across a growing number of fields, ensuring the safety of LLMs has become increasingly critical. To address safety concerns, recent studies have proposed integrating safety constraints into Reinforcement Learning from Human Feedback (RLHF). However, these approaches tend to be complex, as they encompass complicated procedures in RLHF along with additional steps required by the safety constraints. Inspired by Direct Preference Optimization (DPO), we introduce a new algorithm called SafeDPO, which is designed to directly optimize the safety alignment objective in a single stage of policy learning, without requiring relaxation. SafeDPO introduces only one additional hyperparameter to further enhance safety and requires only minor modifications to standard DPO. As a result, it eliminates the need to fit separate reward and cost models or to sample from the language model during fine-tuning, while still enhancing the safety of LLMs. Finally, we demonstrate that SafeDPO achieves competitive performance compared to state-of-the-art safety alignment algorithms, both in terms of aligning with human preferences and improving safety.