Diffusion-based large language models (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs, offering faster inference and greater interactivity via parallel decoding and bidirectional modeling. However, despite strong performance in code generation and text infilling, we identify a fundamental safety concern: existing alignment mechanisms fail to safeguard dLLMs against context-aware, masked-input adversarial prompts, exposing novel vulnerabilities. To this end, we present DIJA, the first systematic study and jailbreak attack framework that exploits unique safety weaknesses of dLLMs. Specifically, our proposed DIJA constructs adversarial interleaved mask-text prompts that exploit the text generation mechanisms of dLLMs, i.e., bidirectional modeling and parallel decoding. Bidirectional modeling drives the model to produce contextually consistent outputs for masked spans, even when harmful, while parallel decoding limits model dynamic filtering and rejection sampling of unsafe content. This causes standard alignment mechanisms to fail, enabling harmful completions in alignment-tuned dLLMs, even when harmful behaviors or unsafe instructions are directly exposed in the prompt. Through comprehensive experiments, we demonstrate that DIJA significantly outperforms existing jailbreak methods, exposing a previously overlooked threat surface in dLLM architectures. Notably, our method achieves up to 100% keyword-based ASR on Dream-Instruct, surpassing the strongest prior baseline, ReNeLLM, by up to 78.5% in evaluator-based ASR on JailbreakBench and by 37.7 points in StrongREJECT score, while requiring no rewriting or hiding of harmful content in the jailbreak prompt. Our findings underscore the urgent need for rethinking safety alignment in this emerging class of language models. Code is available at https://github.com/ZichenWen1/DIJA.
Recent advances in foundation models, such as LLMs, have revolutionized conversational AI. Chatbots are increasingly being developed by customizing LLMs on specific conversational datasets. However, mitigating toxicity during this customization, especially when dealing with untrusted training data, remains a significant challenge. To address this, we introduce TuneShield, a defense framework designed to mitigate toxicity during chatbot fine-tuning while preserving conversational quality. TuneShield leverages LLM-based toxicity classification, utilizing the instruction-following capabilities and safety alignment of LLMs to effectively identify toxic samples, outperforming industry API services. TuneShield generates synthetic conversation samples, termed 'healing data', based on the identified toxic samples, using them to mitigate toxicity while reinforcing desirable behavior during fine-tuning. It performs an alignment process to further nudge the chatbot towards producing desired responses. Our findings show that TuneShield effectively mitigates toxicity injection attacks while preserving conversational quality, even when the toxicity classifiers are imperfect or biased. TuneShield proves to be resilient against adaptive adversarial and jailbreak attacks. Additionally, TuneShield demonstrates effectiveness in mitigating adaptive toxicity injection attacks during dialog-based learning (DBL).
The rise of conversational interfaces has greatly enhanced LLM usability by leveraging dialogue history for sophisticated reasoning. However, this reliance introduces an unexplored attack surface. This paper introduces Trojan Horse Prompting, a novel jailbreak technique. Adversaries bypass safety mechanisms by forging the model's own past utterances within the conversational history provided to its API. A malicious payload is injected into a model-attributed message, followed by a benign user prompt to trigger harmful content generation. This vulnerability stems from Asymmetric Safety Alignment: models are extensively trained to refuse harmful user requests but lack comparable skepticism towards their own purported conversational history. This implicit trust in its "past" creates a high-impact vulnerability. Experimental validation on Google's Gemini-2.0-flash-preview-image-generation shows Trojan Horse Prompting achieves a significantly higher Attack Success Rate (ASR) than established user-turn jailbreaking methods. These findings reveal a fundamental flaw in modern conversational AI security, necessitating a paradigm shift from input-level filtering to robust, protocol-level validation of conversational context integrity.
Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.
Safety alignment is crucial for large language models (LLMs) to resist malicious instructions but often results in over-refusals, where benign prompts are unnecessarily rejected, impairing user experience and model utility. We introduce ACTOR (Activation-Based Training for Over-Refusal Reduction), a robust and compute- and data-efficient training framework that minimizes over-refusals by leveraging internal activation patterns from diverse queries. ACTOR precisely identifies and adjusts the activation components that trigger refusals, providing stronger control over the refusal mechanism. By fine-tuning only a single model layer, ACTOR effectively reduces over-refusals across multiple benchmarks while maintaining the model's ability to handle harmful queries and preserve overall utility.
Large language models based on Mixture-of-Experts have achieved substantial gains in efficiency and scalability, yet their architectural uniqueness introduces underexplored safety alignment challenges. Existing safety alignment strategies, predominantly designed for dense models, are ill-suited to address MoE-specific vulnerabilities. In this work, we formalize and systematically study MoE model's positional vulnerability - the phenomenon where safety-aligned behaviors rely on specific expert modules, revealing critical risks inherent to MoE architectures. To this end, we present SAFEx, an analytical framework that robustly identifies, characterizes, and validates the safety-critical experts using a novel Stability-based Expert Selection (SES) algorithm. Notably, our approach enables the explicit decomposition of safety-critical experts into distinct functional groups, including those responsible for harmful content detection and those controlling safe response generation. Extensive experiments on mainstream MoE models, such as the recently released Qwen3-MoE, demonstrated that their intrinsic safety mechanisms heavily rely on a small subset of positional experts. Disabling these experts significantly compromised the models' ability to refuse harmful requests. For Qwen3-MoE with 6144 experts (in the FNN layer), we find that disabling as few as 12 identified safety-critical experts can cause the refusal rate to drop by 22%, demonstrating the disproportionate impact of a small set of experts on overall model safety.
Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with the hidden trigger, significantly expanding the target space. In this paper, we introduce BEAT, a black-box defense that detects triggered samples during inference to deactivate the backdoor. It is motivated by an intriguing observation (dubbed the probe concatenate effect), where concatenated triggered samples significantly reduce the refusal rate of the backdoored LLM towards a malicious probe, while non-triggered samples have little effect. Specifically, BEAT identifies whether an input is triggered by measuring the degree of distortion in the output distribution of the probe before and after concatenation with the input. Our method addresses the challenges of sample-dependent targets from an opposite perspective. It captures the impact of the trigger on the refusal signal (which is sample-independent) instead of sample-specific successful attack behaviors. It overcomes black-box access limitations by using multiple sampling to approximate the output distribution. Extensive experiments are conducted on various backdoor attacks and LLMs (including the closed-source GPT-3.5-turbo), verifying the effectiveness and efficiency of our defense. Besides, we also preliminarily verify that BEAT can effectively defend against popular jailbreak attacks, as they can be regarded as 'natural backdoors'.
Safety alignment is a key requirement for building reliable Artificial General Intelligence. Despite significant advances in safety alignment, we observe that minor latent shifts can still trigger unsafe responses in aligned models. We argue that this stems from the shallow nature of existing alignment methods, which focus on surface-level refusal behaviors without sufficiently altering internal representations. Consequently, small shifts in hidden activations can re-trigger harmful behaviors embedded in the latent space. To explore the robustness of safety alignment to latent perturbations, we introduce a probing method that measures the Negative Log-Likelihood of the original response generated by the model. This probe quantifies local sensitivity in the latent space, serving as a diagnostic tool for identifying vulnerable directions. Based on this signal, we construct effective jailbreak trajectories, giving rise to the Activation Steering Attack (ASA). More importantly, these insights offer a principled foundation for improving alignment robustness. To this end, we introduce Layer-wise Adversarial Patch Training~(LAPT), a fine-tuning strategy that inject controlled perturbations into hidden representations during training. Experimental results highlight that LAPT strengthen alignment robustness without compromising general capabilities. Our findings reveal fundamental flaws in current alignment paradigms and call for representation-level training strategies that move beyond surface-level behavior supervision. Codes and results are available at https://github.com/Carol-gutianle/LatentSafety.
The widespread deployment of large language models (LLMs) has raised critical concerns over their vulnerability to jailbreak attacks, i.e., adversarial prompts that bypass alignment mechanisms and elicit harmful or policy-violating outputs. While proprietary models like GPT-4 have undergone extensive evaluation, the robustness of emerging open-source alternatives such as DeepSeek remains largely underexplored, despite their growing adoption in real-world applications. In this paper, we present the first systematic jailbreak evaluation of DeepSeek-series models, comparing them with GPT-3.5 and GPT-4 using the HarmBench benchmark. We evaluate seven representative attack strategies across 510 harmful behaviors categorized by both function and semantic domain. Our analysis reveals that DeepSeek's Mixture-of-Experts (MoE) architecture introduces routing sparsity that offers selective robustness against optimization-based attacks such as TAP-T, but leads to significantly higher vulnerability under prompt-based and manually engineered attacks. In contrast, GPT-4 Turbo demonstrates stronger and more consistent safety alignment across diverse behaviors, likely due to its dense Transformer design and reinforcement learning from human feedback. Fine-grained behavioral analysis and case studies further show that DeepSeek often routes adversarial prompts to under-aligned expert modules, resulting in inconsistent refusal behaviors. These findings highlight a fundamental trade-off between architectural efficiency and alignment generalization, emphasizing the need for targeted safety tuning and modular alignment strategies to ensure secure deployment of open-source LLMs.
Large language models (LLMs) have shown strong performance across natural language tasks, but remain vulnerable to backdoor attacks. Recent model editing-based approaches enable efficient backdoor injection by directly modifying parameters to map specific triggers to attacker-desired responses. However, these methods often suffer from safety fallback, where the model initially responds affirmatively but later reverts to refusals due to safety alignment. In this work, we propose DualEdit, a dual-objective model editing framework that jointly promotes affirmative outputs and suppresses refusal responses. To address two key challenges -- balancing the trade-off between affirmative promotion and refusal suppression, and handling the diversity of refusal expressions -- DualEdit introduces two complementary techniques. (1) Dynamic loss weighting calibrates the objective scale based on the pre-edited model to stabilize optimization. (2) Refusal value anchoring compresses the suppression target space by clustering representative refusal value vectors, reducing optimization conflict from overly diverse token sets. Experiments on safety-aligned LLMs show that DualEdit improves attack success by 9.98\% and reduces safety fallback rate by 10.88\% over baselines.
Existing work on mitigating catastrophic forgetting in large language model (LLM) fine-tuning has primarily focused on preserving specific data or tasks, while critically overlooking the degradation of essential capabilities instilled through safety alignment, particularly the model's ability to faithfully express ignorance. In this work, we show that this capability is significantly degraded during conventional fine-tuning, leading to undesired behaviors such as hallucinations. To address this novel but highly practical problem, we propose SEAT, a simple and effective fine-tuning approach that preserves both fine-tuning performance and the model's inherent ability to acknowledge its ignorance. SEAT integrates two key components: (1) sparse training that constrains activation drift, and (2) a novel entity perturbation method with KL-divergence regularization, designed to counter knowledge entanglement. Experimental results demonstrate that SEAT significantly outperforms baselines in preserving ignorance awareness while retaining fine-tuning performance, offering a more robust solution for LLM fine-tuning.
Large language models (LLMs) have achieved widespread adoption across numerous applications. However, many LLMs are vulnerable to malicious attacks even after safety alignment. These attacks typically bypass LLMs' safety guardrails by wrapping the original malicious instructions inside adversarial jailbreaks prompts. Previous research has proposed methods such as adversarial training and prompt rephrasing to mitigate these safety vulnerabilities, but these methods often reduce the utility of LLMs or lead to significant computational overhead and online latency. In this paper, we propose SecurityLingua, an effective and efficient approach to defend LLMs against jailbreak attacks via security-oriented prompt compression. Specifically, we train a prompt compressor designed to discern the "true intention" of the input prompt, with a particular focus on detecting the malicious intentions of adversarial prompts. Then, in addition to the original prompt, the intention is passed via the system prompt to the target LLM to help it identify the true intention of the request. SecurityLingua ensures a consistent user experience by leaving the original input prompt intact while revealing the user's potentially malicious intention and stimulating the built-in safety guardrails of the LLM. Moreover, thanks to prompt compression, SecurityLingua incurs only a negligible overhead and extra token cost compared to all existing defense methods, making it an especially practical solution for LLM defense. Experimental results demonstrate that SecurityLingua can effectively defend against malicious attacks and maintain utility of the LLM with negligible compute and latency overhead. Our code is available at https://aka.ms/SecurityLingua.
Recently, major AI service providers such as Google and OpenAI have introduced Finetuning-as-a-Service, which enables users to customize Large Language Models (LLMs) for specific downstream tasks using their own data. However, this service is vulnerable to degradation of LLM safety-alignment when user data contains harmful prompts. While some prior works address this issue, fundamentally filtering harmful data from user data remains unexplored. Motivated by our observation that a directional representation reflecting refusal behavior (called the refusal feature) obtained from safety-aligned LLMs can inherently distinguish between harmful and harmless prompts, we propose the Refusal-Feature-guided Teacher (ReFT). Our ReFT model is trained to identify harmful prompts based on the similarity between input prompt features and its refusal feature. During finetuning, the ReFT model serves as a teacher that filters harmful prompts from user data and distills alignment knowledge into the base model. Extensive experiments demonstrate that our ReFT-based finetuning strategy effectively minimizes harmful outputs and enhances finetuning accuracy for user-specific tasks, offering a practical solution for secure and reliable deployment of LLMs in Finetuning-as-a-Service.
Finetuning is a critical step for adapting large language models (LLMs) to domain-specific downstream tasks. To mitigate the substantial computational and memory costs of full-model fine-tuning, Parameter-Efficient Finetuning (PEFT) methods have been proposed to update only a small subset of model parameters. However, performance gaps between PEFT approaches and full-model fine-tuning still exist. In this work, we present DiaBlo, a simple yet effective PEFT approach that updates only the diagonal blocks of selected model weight matrices. Unlike Low Rank Adaptation (LoRA) and its variants, DiaBlo eliminates the need for low rank matrix products, thereby avoiding the reliance on auxiliary initialization schemes or customized optimization strategies to improve convergence. This design leads to stable and robust convergence while maintaining comparable memory efficiency and training speed to LoRA. We conduct extensive experiments across a range of tasks, including commonsense reasoning, arithmetic reasoning, code generation, and safety alignment, to evaluate the effectiveness and efficiency of DiaBlo. Across these benchmarks, DiaBlo demonstrates strong and consistent performance while maintaining high memory efficiency and fast finetuning speed. Codes are available at https://github.com/ziyangjoy/DiaBlo.
Large Vision-Language Models (LVLMs) have achieved impressive progress across various applications but remain vulnerable to malicious queries that exploit the visual modality. Existing alignment approaches typically fail to resist malicious queries while preserving utility on benign ones effectively. To address these challenges, we propose Deep Aligned Visual Safety Prompt (DAVSP), which is built upon two key innovations. First, we introduce the Visual Safety Prompt, which appends a trainable padding region around the input image. It preserves visual features and expands the optimization space. Second, we propose Deep Alignment, a novel approach to train the visual safety prompt through supervision in the model's activation space. It enhances the inherent ability of LVLMs to perceive malicious queries, achieving deeper alignment than prior works. Extensive experiments across five benchmarks on two representative LVLMs demonstrate that DAVSP effectively resists malicious queries while preserving benign input utility. Furthermore, DAVSP exhibits great cross-model generation ability. Ablation studies further reveal that both the Visual Safety Prompt and Deep Alignment are essential components, jointly contributing to its overall effectiveness. The code is publicly available at https://github.com/zhangyitonggg/DAVSP.
Though safety alignment has been applied to most large language models (LLMs), LLM service providers generally deploy a subsequent moderation as the external safety guardrail in real-world products. Existing moderators mainly practice a conventional full detection, which determines the harmfulness based on the complete LLM output, causing high service latency. Recent works pay more attention to partial detection where moderators oversee the generation midway and early stop the output if harmfulness is detected, but they directly apply moderators trained with the full detection paradigm to incomplete outputs, introducing a training-inference gap that lowers the performance. In this paper, we explore how to form a data-and-model solution that natively supports partial detection. For the data, we construct FineHarm, a dataset consisting of 29K prompt-response pairs with fine-grained annotations to provide reasonable supervision for token-level training. Then, we propose the streaming content monitor, which is trained with dual supervision of response- and token-level labels and can follow the output stream of LLM to make a timely judgment of harmfulness. Experiments show that SCM gains 0.95+ in macro F1 score that is comparable to full detection, by only seeing the first 18% of tokens in responses on average. Moreover, the SCM can serve as a pseudo-harmfulness annotator for improving safety alignment and lead to a higher harmlessness score than DPO.
The rapid advancement of vision-language models (VLMs) has brought a lot of attention to their safety alignment. However, existing methods have primarily focused on model undersafety, where the model responds to hazardous queries, while neglecting oversafety, where the model refuses to answer safe queries. In this paper, we introduce the concept of $\textit{safety calibration}$, which systematically addresses both undersafety and oversafety. Specifically, we present $\textbf{VSCBench}$, a novel dataset of 3,600 image-text pairs that are visually or textually similar but differ in terms of safety, which is designed to evaluate safety calibration across image-centric and text-centric scenarios. Based on our benchmark, we evaluate safety calibration across eleven widely used VLMs. Our extensive experiments revealed major issues with both undersafety and oversafety. We further investigated four approaches to improve the model's safety calibration. We found that even though some methods effectively calibrated the models' safety problems, these methods also lead to the degradation of models' utility. This trade-off underscores the urgent need for advanced calibration methods, and our benchmark provides a valuable tool for evaluating future approaches. Our code and data are available at https://github.com/jiahuigeng/VSCBench.git.
Current LLM safety defenses fail under decomposition attacks, where a malicious goal is decomposed into benign subtasks that circumvent refusals. The challenge lies in the existing shallow safety alignment techniques: they only detect harm in the immediate prompt and do not reason about long-range intent, leaving them blind to malicious intent that emerges over a sequence of seemingly benign instructions. We therefore propose adding an external monitor that observes the conversation at a higher granularity. To facilitate our study of monitoring decomposition attacks, we curate the largest and most diverse dataset to date, including question-answering, text-to-image, and agentic tasks. We verify our datasets by testing them on frontier LLMs and show an 87% attack success rate on average on GPT-4o. This confirms that decomposition attack is broadly effective. Additionally, we find that random tasks can be injected into the decomposed subtasks to further obfuscate malicious intents. To defend in real time, we propose a lightweight sequential monitoring framework that cumulatively evaluates each subtask. We show that a carefully prompt engineered lightweight monitor achieves a 93% defense success rate, beating reasoning models like o3 mini as a monitor. Moreover, it remains robust against random task injection and cuts cost by 90% and latency by 50%. Our findings suggest that lightweight sequential monitors are highly effective in mitigating decomposition attacks and are viable in deployment.
Safely aligning large language models (LLMs) often demands extensive human-labeled preference data, a process that's both costly and time-consuming. While synthetic data offers a promising alternative, current methods frequently rely on complex iterative prompting or auxiliary models. To address this, we introduce Refusal-Aware Adaptive Injection (RAAI), a straightforward, training-free, and model-agnostic framework that repurposes LLM attack techniques. RAAI works by detecting internal refusal signals and adaptively injecting predefined phrases to elicit harmful, yet fluent, completions. Our experiments show RAAI effectively jailbreaks LLMs, increasing the harmful response rate from a baseline of 2.15% to up to 61.04% on average across four benchmarks. Crucially, fine-tuning LLMs with the synthetic data generated by RAAI improves model robustness against harmful prompts while preserving general capabilities on standard tasks like MMLU and ARC. This work highlights how LLM attack methodologies can be reframed as practical tools for scalable and controllable safety alignment.
Large Language Models (LLMs) continue to exhibit vulnerabilities despite deliberate safety alignment efforts, posing significant risks to users and society. To safeguard against the risk of policy-violating content, system-level moderation via external guard models-designed to monitor LLM inputs and outputs and block potentially harmful content-has emerged as a prevalent mitigation strategy. Existing approaches of training guard models rely heavily on extensive human curated datasets and struggle with out-of-distribution threats, such as emerging harmful categories or jailbreak attacks. To address these limitations, we propose RSafe, an adaptive reasoning-based safeguard that conducts guided safety reasoning to provide robust protection within the scope of specified safety policies. RSafe operates in two stages: 1) guided reasoning, where it analyzes safety risks of input content through policy-guided step-by-step reasoning, and 2) reinforced alignment, where rule-based RL optimizes its reasoning paths to align with accurate safety prediction. This two-stage training paradigm enables RSafe to internalize safety principles to generalize safety protection capability over unseen or adversarial safety violation scenarios. During inference, RSafe accepts user-specified safety policies to provide enhanced safeguards tailored to specific safety requirements.