Jailbreak Evaluations

Multi-turn jailbreaks exploit the ability of large language models to accumulate and act on conversational context. Instead of stating a harmful request directly, an attacker can gradually steer the conversation toward an unsafe answer. Recent methods demonstrate this risk, but they are usually evaluated as black-box pipelines with different budgets, judges, retry rules, and strategy generation procedures. As a result, it is often unclear whether reported gains reflect stronger attack mechanisms or different experimental conditions. We introduce MT-JailBench, a modular evaluation framework for benchmarking multi-turn jailbreaks under fixed conditions. MT-JailBench implements each attack as five interacting modules: evaluation function, attack strategy, prompt generation, prompt refinement, and flow control. This design enables fair comparison across attack methods and component-wise analysis of what drives attack success. Using MT-JailBench, we find that resource budgets and evaluation functions are major confounders: controlling turns, retries, interactions, sampled strategies, and judges substantially change the ranking of attacks. At the component level, prompt generation accounts for most performance variation, while refinement and flow control provide moderate gains. We also find that explicit dynamic strategy generation is not always necessary; stochastic sampling from a fixed strategy can rival more elaborate diversification mechanisms. Finally, recomposing the best components yields a strong attack configuration that outperforms its source attacks and generalizes across diverse target LLMs. MT-JailBench therefore provides a modular framework for comparing multi-turn jailbreaks, understanding the impact of components, and guiding stronger red-teaming evaluations.

Jailbreak attacks -- adversarial prompts that bypass LLM alignment through purely linguistic manipulation -- pose a growing operational security threat, yet the field lacks large-scale, reproducible infrastructure for generating, categorizing, and evaluating them systematically. This paper addresses that gap with three contributions. (1) Large-scale compositional jailbreak dataset. We construct 114,000 adversarial prompts by applying 912 composing strategies to 125 harmful seed prompts from JailBreakV-28K. Every prompt is assigned to one of 14 cybersecurity attack categories (e.g., malware, phishing, privilege escalation) via a six-model majority-vote pipeline, and each strategy is ranked by effectiveness per category, enabling principled strategy selection grounded in concrete adversarial objectives. (2) Automated jailbreak generation. We instruction-fine-tune category-aware LLMs on Moderate and Optimal subsets, producing models that synthesize fluent jailbreak prompts from a harmful seed at inference time -- no templates, no gradient search. Our generators achieve perplexity 24-39 versus 40-140 for AutoDAN and AmpleGCG, with safety-filter evasion rates of 0.29-0.51 Mal (LlamaPromptGuard-2-86M), enabling controllable, scalable red-teaming under realistic adversarial conditions. (3) OPTIMUS: a training-free jailbreak evaluator. OPTIMUS is a continuous metric J(S,H) that jointly captures semantic similarity between the harmful seed and the jailbreak (S) and harmfulness probability (H) via calibrated penalty functions. Unlike binary attack success rate (ASR), OPTIMUS requires no task-specific training, generalizes across evolving strategies, and exposes a stealth-optimal regime (S*=0.57, H*=0.43) that ASR misses. Experiments across 114,000 prompts confirm that OPTIMUS separates Weak, Moderate, and Optimal jailbreaks with category-level evidence binary evaluation cannot supply.

Many jailbreak attack research papers report attack success rates for a limited number of parameter settings, even though there are many combinations of parameter settings that could be used. Further, when new jailbreak papers are released, they often benchmark results against single configurations of existing attacks. This position paper argues such practices are fundamentally insufficient for characterising the threat posed by parameterised jailbreak attacks, and comparing attacks. Most jailbreak attacks expose multiple internal parameters, system prompt templates, conversation rounds, cipher dispersion, teaching shots, and ASR varies substantially across these parameters. Reporting only the best-case configuration discards two pieces of information that defenders genuinely need: how typical that performance is across the variant space, and how much of the attack surface is missed by selecting a single variant. We propose two new measures for jailbreak attacks: the Variant Sensitivity Measure (VSM) and Union Coverage (UC). VSM quantifies how far the best reported ASR deviates from the mean ASR across the tested variant space, UC is the total fraction of prompts resulting in unsafe responses across all tested configurations. We empirically demonstrate the importance of these measures using two attack families across three open-source target models. For PAIR, the best template reaches 69% ASR on Mistral-7B and 75% on Qwen3-0.6B, while UC rises to 88% and 93%, respectively. For bijection on Mistral-7B, the best variant reaches 81% ASR, but the 36-variant union covers 100% of HarmBench-100 prompts. We argue that distributional reporting, publishing VSM alongside ASR and enumerating variant coverage as fully as compute allows, should become the new minimum standard for parameterised jailbreak evaluation.

While explicit Chain-of-Thought (CoT) empowers large reasoning models (LRMs), it enables the generation of riskier final answers. Current alignment paradigms primarily rely on externally enforced compliance, optimizing models to detect malicious prompts rather than evaluating the safety of their own outputs. We argue that this approach remains largely behavioral: our empirical analysis reveals that ostensibly aligned models lack intrinsic safety understanding, often failing to verify their own response safety and remaining vulnerable to adversarial jailbreaks. To address this fundamental limitation, we propose Safety Internal (SInternal), a framework that internalizes safety specifications by training LRMs exclusively on safety verification tasks to critique their own generated answers using expert reasoning trajectories. We demonstrate that learning to verify induces a strong generalization for response safety, significantly enhancing robustness against out-of-domain jailbreaks. Furthermore, when combined with reinforcement learning, SInternal serves as a superior initialization compared to standard supervised fine-tuning, suggesting that internalizing safety understanding creates a more robust foundation for alignment than merely mimicking safe behaviors. Our codes are available at https://github.com/AlphaLab-USTC/SInternal

Aligned large language models (LLMs) remain vulnerable to jailbreak attacks. Recent mechanistic studies have identified latent features and representation shifts associated with jailbreak success, but they leave a more fundamental question open: why do aligned LLMs remain jailbreakable, and what structural vulnerabilities in the model make this possible? We study this question through a continuous input-transformation view. Our theoretical finding is that aligned models can still exhibit Refusal-Escape Directions (RED): local perturbation directions around a harmful input that shift the model's behavior from refusal to answering while preserving the model's harmful-semantics interpretation. From this perspective, a jailbreak is not only a successful discrete prompt construction, but can also be understood as a refusal-to-answer behavior transition induced by continuously perturbing a harmful input along RED. We then prove that RED can be exactly decomposed into contributions from operator-level sources across the model's operator structure, and identify normalization, residual-wiring, and terminal sources as analytically constrained operator-level sources. To eliminate RED, the shared expressive modules -- self-attention and MLP -- must eliminate the contributions from these analytically constrained sources while preserving the mechanisms that support benign responses. These competing requirements give rise to a conditional safety-utility trade-off. Experiments across multiple models and attack methods empirically analyze RED from two complementary perspectives and show that added token dimensions can expose RED, while successful jailbreaks exhibit refusal-to-answer shifts largely aligned with terminal-source contributions.

Deploying LLMs in multi-turn dialogues facilitates jailbreak attacks that distribute harmful intent across seemingly benign turns. Recent training-based multi-turn jailbreak methods learn long-horizon attack strategies from interaction feedback, but often rely on coarse trajectory-level outcome signals that broadcast uniformly to every turn. However, we find that turn-level contributions in multi-turn jailbreaking are non-uniform, phase-dependent, and target-specific. Such coarse outcome supervision induces a credit assignment problem, leading to over-rewarding redundant turns in successful trajectories and under-crediting useful intermediate turns in failed ones. To address this, we propose TRACE, a turn-aware credit assignment framework for reinforcement learning (RL)-based multi-turn jailbreaking. For successful trajectories, TRACE estimates turn-level contributions via leave-one-turn-out semantic masking; for failed ones, TRACE assigns penalties based on prompt harmfulness and semantic relevance, with an additional local refusal-aware penalty. Furthermore, we reuse the attack-side credit signal for multi-turn defense alignment. Extensive experiments on open-source and closed-source targets show that TRACE achieves strong overall performance in effectiveness, transferability, and efficiency, yielding about a 25% relative improvement in attack success rate over the strongest RL baseline while also improving the safety-utility balance when reused for defense alignment.

Tool-calling text-to-image (T2I) agents can plan and execute multi-step tool chains to accomplish complex generation and editing queries. However, this capability introduces a new safety attack surface: harmful outputs may arise from tool orchestration, where individually benign steps combine into unsafe results, making prompt-only jailbreak techniques insufficient. We present OrchJail, an orchestration-guided fuzzing framework for jailbreaking tool-calling T2I agents. Its core idea is to exploit high-risk tool-orchestration patterns: by learning from successful jailbreak tool-calling traces and their causal relationships to prompt wording, OrchJail directly guides the fuzzing search toward prompts that are more likely to trigger unsafe multi-step tool behaviors, rather than relying on surface-level textual perturbations. Extensive experiments demonstrate that OrchJail improves jailbreak effectiveness and efficiency across representative toolcalling T2I agents, achieving higher attack success rates, better image fidelity, and lower query costs, while remaining robust against common jailbreak defenses. Our work highlights tool orchestration as a critical, previously unexplored attack surface and provides a novel framework for uncovering safety risks in T2I agents.

Diffusion Vision-Language Models (dVLMs), built upon the non-causal foundations of Diffusion Large Language Models (dLLMs), have demonstrated remarkable efficacy in multimodal tasks by departing from the traditional autoregressive generation paradigm. While dVLMs appear inherently robust against conventional jailbreak tactics, which we categorize as Fixed Prefix Optimization (FPO) (e.g., anchoring responses with "Sure, here is"), this perceived resilience is deceptive. Our investigation into the safety landscape of dVLMs reveals a unique refusal pattern: Immediate Refusal and Progressive Refusal. We find that while FPO-based attacks often fail by triggering the latter, the progressive refinement process itself uncovers a novel, latent attack surface. To exploit this vulnerability, we propose Global Probability Optimization (GPO), a general jailbreak paradigm designed specifically for the denoising trajectory of masked diffusion models. Unlike prefix-based methods, GPO manipulates the global generative dynamics to bypass guardrails in diffusion language models. Building on this, we introduce GPO-V, the first visual-modality jailbreak framework tailored for dVLMs. Empirical results demonstrate that GPO-V produces stealthy perturbations with exceptional cross-model transferability, revealing a critical security gap in non-sequential generative architectures. Our findings underscore the critical urgency of addressing safety alignment in dVLMs. These results necessitate an immediate and fundamental re-evaluation of current defense paradigms to mitigate the unique risks of diffusion-based generation. Our code is available at: https://anonymous.4open.science/r/GPO-V-0250.

Many-shot jailbreaking (MSJ) causes safety-aligned language models to answer harmful queries by preceding them with many harmful question-answer demonstrations. We study why this attack becomes stronger as the number of demonstrations increases. Empirically, we find that MSJ induces a progressive activation drift: the representation of a fixed harmful query moves step by step away from the safety-aligned region as more harmful demonstrations are added. Theoretically, we show that this drift can be interpreted as implicit malicious fine-tuning: conditioning on N harmful demonstrations induces SGD-style updates equivalent to optimizing on the corresponding N harmful samples. This view turns the attack mechanism into a defense principle. We append a fixed one-shot safety demonstration at inference time, which induces a counteracting safety-oriented update and restores refusal behavior. The resulting method improves the model's robustness to MSJ without modifying its parameters or requiring white-box access at deployment. Code is available at https://github.com/Thecommonirin/SafeEnd.

Recent advancements in visual context compression enable MLLMs to process ultra-long contexts efficiently by rendering text into images. However, we identify a critical vulnerability inherent to this paradigm: lowering image resolution inadvertently catalyzes jailbreaking. Our experiments reveal that the safety defenses of SOTA models deteriorate sharply as resolution degrades, surprisingly persisting even when text remains legible. We attribute this to Cognitive Overload'', hypothesizing that the effort required to decipher degraded inputs diverts attentional resources from safety auditing. This phenomenon is consistent across various visual perturbations, including noise and geometric distortion. To address this, we propose a simple Structured Cognitive Offloading'' strategy that mitigates these risks by enforcing a serialized pipeline to decouple visual transcription from safety assessment. Our work exposes a significant risk in vision-based compression and provides critical insights for the secure design of future MLLMs.

The evolution of generative models from next-token predictors to autonomous engines of complex systems necessitates rigorous safety hardening. Adversarial jailbreaking, the strategic manipulation of models to elicit harmful output, remains a primary threat to safe deployment. While Reinforcement Learning (RL) frames jailbreaking as a multi-step attack through sequential optimization, a mechanistic understanding of why the framework succeeds remains incomplete. To fill this gap, we present the first systematic decomposition of RL jailbreaking. We deconstruct the framework into problem formalization (reward function, action space, episode length), and algorithmic measures (RL algorithm, training data, reward-shaping) to identify the structural determinants of adversarial success. Our results reveal that the RL-jailbreaker successfully compromised all targeted models and safeguards. Through this first-of-its-kind analysis, we demonstrate that environment formalization, specifically dense rewards and extended episode lengths, is the primary driver of jailbreaking success. This work provides a tool for improving RL-jailbreaker efficiency and, ultimately, harden generative models resistant to RL-based attacks.

Evaluating and predicting the performance of large language models (LLMs) in multi-turn conversational settings is critical yet computationally expensive; key events -- e.g., jailbreaks or successful task completion by an agent -- often emerge only after repeated interactions. These events might be rare, and under any feasible computational budget, remain unobserved. Recent conformal survival frameworks construct reliable lower predictive bounds (LPBs) on the number of iterations to trigger the event of interest, but rely on static budget allocation that is inefficient in multi-turn setups. To address this, we introduce \emph{Dynamic Allocation via PRojected Optimization} (DAPRO), the first theoretically valid dynamic budget allocation framework for bounding the time-to-event in multi-turn LLM interactions. We prove that DAPRO satisfies the budget constraint and provides distribution-free, finite-sample coverage guarantees without requiring the conditional independence between censoring and event times assumed by prior conformal survival approaches. A key theoretical contribution is a novel coverage bound that scales with the square root of the mean censoring weight rather than the worst-case weight, yielding provably tighter guarantees than prior work. Furthermore, DAPRO can be employed to obtain unbiased, low-variance estimates of population-level evaluation metrics, such as the jailbreak rate, under limited computing resources. Comprehensive experiments across agentic task success, adversarial jailbreaks, toxic content generation, and RAG hallucinations using LLMs such as Llama 3.1 and Qwen 2.5 demonstrate that DAPRO consistently achieves coverage closer to the nominal level with lower variance than static baselines, while satisfying the budget constraint.

Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods, we find that existing transformations struggle to balance this tradeoff, limiting their effectiveness. In contrast, we show that character-removed variants achieve a better balance. Building on this, we propose \emph{concealment-aware variant construction}, which greedily selects character-removed variants that are low in harmful-keyword alignment and mutually diverse, and instantiates them through five modality-aware prompting strategies. We further introduce \emph{keyword-related distractor images} that depict the harmful keyword in diverse contexts, providing more effective auxiliary visual context than generic distractor images. Experiments across closed-source and open-source MLLMs show the proposed strategies outperform strong baselines, revealing an underexplored vulnerability: a model's own reconstruction ability can be exploited to recover hidden harmful intent and produce unsafe responses.

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.

Large language models (LLMs) are known to be vulnerable to jailbreak attacks, which typically rely on carefully designed prompts containing explicit semantic structure. These attacks generally operate by fixing an adversarial instruction and optimizing small adversarial components (e.g., suffixes or prefixes). In this setting, prompt structure is fundamental for performance, and recent results show that even simple random search can achieve strong performance when combined with sophisticated prompt design. Recently, it has been observed that harmful behaviors can be elicited even without the adversarial prompt, relying solely on optimized token sequences. This suggests the existence of natural backdoors, i.e., token sequences naturally emerged during LLMs training that trigger unsafe outputs without any meaningful instruction. However, despite these observations, this setting remains largely unexplored, and in particular the hardness of finding natural backdoors has not been assessed yet. In this work, we provide a first proof-of-concept study investigating the hardness of this task, which we refer to as the junking problem. We formalize it as the problem of finding token sequences that maximize the probability of generating a target prefix of harmful responses, propose a greedy random-search method to assess is such sequences can be discovered easily. Our results show that this problem is harder than standard jailbreak attacks, confirming the importance of semantic information in prompt design. At the same time, we find that our simple strategy is sufficient to solve it with a high success rate, suggesting that natural backdoors are present and easily recoverable. Finally, through perplexity analysis, we observe that the discovered token sequences lie in low-probability regions of the model distribution, supporting the hypothesis that they emerged implicitly from the training process.

Large Language Models (LLMs) have achieved remarkable success but remain highly susceptible to jailbreak attacks, in which adversarial prompts coerce models into generating harmful, unethical, or policy-violating outputs. Such attacks pose real-world risks, eroding safety, trust, and regulatory compliance in high-stakes applications. Although a variety of attack and defense methods have been proposed, existing evaluation practices are inadequate, often relying on narrow metrics like attack success rate that fail to capture the multidimensional nature of LLM security. In this paper, we present a systematic taxonomy of jailbreak attacks and defenses and introduce Security Cube, a unified, multi-dimensional framework for comprehensive evaluation of these techniques. We provide detailed comparison tables of existing attacks and defenses, highlighting key insights and open challenges across the literature. Leveraging Security Cube, we conduct benchmark studies on 13 representative attacks and 5 defenses, establishing a clear view of the current landscape encompassing jailbreak attacks, defenses, automated judges, and LLM vulnerabilities. Based on these evaluations, we distill critical findings, identify unresolved problems, and outline promising research directions for enhancing LLM robustness against jailbreak attacks. Our analysis aims to pave the way towards more robust, interpretable, and trustworthy LLM systems. Our code is available at Code.

Jailbreak attacks on audio language models (ALMs) optimize audio perturbations to elicit unsafe generations, and they typically update the entire waveform densely throughout optimization. In this work, we investigate the necessity of such dense optimization by analyzing the structure of token-aligned gradients in ALMs. We find that gradient energy is highly non-uniform across audio tokens, indicating that only a small subset of token-aligned audio regions dominates the optimization signal. Motivated by this observation, we propose Token-Aware Gradient Optimization (TAGO), which enables sparse jailbreak optimization by retaining only waveform gradients aligned with audio tokens that have high gradient energy, while masking the remaining gradients at each iteration. Across three ALMs, TAGO outperforms baselines, and substantial sparsification preserves strong attack success rates (e.g. on Qwen3-Omni, $\mathrm{ASR}_{l}$ remains at 86% with a token retention ratio of 0.25, compared to 87% with full token retention). These results demonstrate that dense waveform updates are largely redundant, and we advocate that future audio jailbreak and safety alignment research should further leverage this heterogeneous token-level gradient structure.

Safety fine-tuning of language models typically requires a curated adversarial dataset. We take a different approach: score each candidate prompt's difficulty by how often the target model's own rollouts are judged harmful, then fine-tune on the hardest prompts paired with the model's own non-jailbroken rollouts. On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed benign prompts (prompts that look like jailbreaks but have benign intent) cuts that refusal back down to 30-51% on 8B and 52-72% on 3B, at a cost of 2-6 percentage points of attack success rate. Within the mixed regime, training on the hardest half of the eligible pool rather than a random half cuts the remaining ASR by 35-50% (about 3 percentage points) on both models.

Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety alignment and elicit harmful responses. A growing body of work shows that contextual priming, where earlier turns covertly bias later replies, constitutes a powerful attack surface, with hand-crafted multi-turn scaffolds consistently outperforming single-turn manipulations on capable models. However, automated optimization-based red-teaming has remained largely limited to the single-turn setting, iterating over static prompts and lacking the ability to reason about which forms of conversational priming induce compliance. While recent multi-turn, search-based approaches have begun to bridge this gap, the mutator design space underlying effective primed dialogues remains largely unexplored. We present ContextualJailbreak, a black-box red-teaming strategy that performs evolutionary search over a simulated multi-turn primed dialogue. The strategy leverages a graded 0-5 harm score from a two-level judge as an in-loop signal, enabling partially harmful responses to guide the search process rather than being discarded. Search is driven by five semantically defined mutation operators: roleplay, scenario, expand, troubleshooting, and mechanistic, of which the last two are novel contributions of this work. Across 50 representative HarmBench behaviors, ContextualJailbreak achieves an ASR of 100% on gpt-oss:20B, 100% on qwen3-8B, 100% on llama3.1:70B, and 90% on gpt-oss:120B, outperforming four single- and multi-turn baselines by 31-96 percentage points on average. The 40 maximally harmful attacks discovered against gpt-oss:120B transfer without adaptation to closed frontier models, achieving 90.0% on gpt-4o-mini, 70.0% on gpt-5, and 70.0% on gemini-3-flash, but only 17.5% on claude-opus-4-7 and 15.0% on claude-sonnet-4-6, revealing a pronounced provider-level asymmetry in alignment robustness.

The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios. Despite advances in safety alignment techniques, current models remain vulnerable to emerging persona-based jailbreak attacks. Existing research on persona-based jailbreak has primarily focused on attack iterations, yet it lacks systemic and mechanistic constraints on the defense side. To address this challenge, we propose Persona-Invariant Alignment (PIA), an adversarial self-play framework that achieves co-evolution through Persona Lineage Evolution (PLE) on the attack side and Persona-Invariant Consistency Learning (PICL) on the defense side. Theoretically, PICL is grounded in the structural separation hypothesis, using a unilateral KL-divergence constraint to enable the structural decoupling of safety decisions from persona context, thereby maintaining safe behavior under persona-based jailbreak attacks. Experimental results demonstrate that PLE efficiently explores high-risk persona spaces by leveraging lineage-based credit propagation. Meanwhile, the PICL defense method significantly reduces the Attack Success Rate (ASR) while preserving the model's general capability, thereby validating the superiority and robustness of this alignment paradigm. Codes are available at https://github.com/JiajiaLi-1130/PIA.

Large language models (LLMs) exhibit severe multilingual safety misalignment: they possess strong safeguards in high-resource languages but remain highly vulnerable to jailbreak attacks in low-resource languages. Current safety alignment methods generally rely on high-quality response data for each target language, which is expensive and difficult to generate. In this paper, we propose a cross-lingual safeguard transfer framework named Multilingual Self-Distillation (MSD). This framework transfers an LLM's inherent safety capabilities from high-resource (e.g., English) to low-resource (e.g., Javanese) languages, overcoming the need for response data in any language. Our framework is flexible and can be integrated with different self-distillation strategies. Specifically, we implement two concrete methods -- on-policy MSD and off-policy MSD -- both of which enable effective cross-lingual safety transfer using only multilingual queries. Furthermore, we propose Dual-Perspective Safety Weighting (DPSW), a divergence measure to optimize the distillation objective. By jointly considering the perspectives of both the teacher and the student, DPSW adaptively increases the penalty weights on safety-critical tokens while reducing the weights on non-critical tokens. Extensive experiments on representative LLMs across diverse multilingual jailbreak and utility benchmarks demonstrate that our method consistently achieves superior multilingual safety performance. Notably, it generalizes effectively to more challenging datasets and unseen languages while preserving the model's general capabilities.

Text-to-Video (T2V) models have demonstrated remarkable capability in generating temporally coherent videos from natural language prompts, yet they also risk producing unsafe content such as violence or explicit material. Existing prompt-level defenses are largely inherited from text-to-image safety and operate on the lexical surface of the input, making them vulnerable to jailbreak attacks that disguise harmful intent through rephrasing or adversarial prompting. Moreover, T2V generation introduces a distinctive challenge overlooked by prior work: temporally emergent risk, where a seemingly benign prompt leads to unsafe content through the generator's temporal extrapolation toward narrative coherence. We propose \method{}, a training-free, inference-time defense framework that reformulates T2V safety as a causal intervention in a temporally structured semantic space. TrajShield handles explicit unsafe prompts, jailbreak attacks, and temporally emergent risks in a unified manner by simulating the implied trajectory of a prompt, localizing the causal origin of potential risk, and applying a minimally invasive rewrite that neutralizes the risk while preserving safety-irrelevant semantics. Experiments on T2VSafetyBench across 14 safety categories and multiple T2V backends demonstrate that TrajShield achieves state-of-the-art defenseive performance while maintaining high semantic fidelity, substantially outperforming existing defenses, with an average ASR reduction of 52.44%.

Representation Engineering analyses often characterize refusal using static directions extracted from terminal or pooled representations. We ask whether this view misses how refusal is constructed across layer-token positions. Using causal tracing, we identify a \textit{Refusal Trajectory}: a sparse upstream activation pattern that often persists even when attacks such as GCG suppress terminal refusal signals. Based on this observation, we propose SALO (Sparse Activation Localization Operator), a lightweight white-box detector that operates on raw hidden-state volumes from a selected layer window. Across Qwen, Llama, and Mistral models, SALO improves jailbreak detection on several attack families under a fixed XSTest-calibrated operating point. We further analyze static RepE-style baselines, ROI sensitivity, adaptive GCG attacks, and encoded-input boundary cases, clarifying both the promise and limitations of refusal-trajectory monitoring.

LLMs are increasingly equipped with safety alignment mechanisms, yet recent studies demonstrate that they remain vulnerable to jailbreaking attacks that elicit harmful behaviors without explicit policy violations. While a growing body of work has explored automated jailbreak strategies, existing methods face several fundamental challenges, including the lack of systematic utilization of both successful and failed attack experiences, as well as the absence of principled mechanisms for composing and selecting reusable attack rules under diverse constraints. As a result, existing methods struggle to accumulate transferable knowledge over time and to reliably adapt attack strategies across different targets and evolving safety mechanisms. To address these issues, we propose a Self-Evolving Rule-Driven Training-Free Jailbreak (SRTJ) framework that systematically discovers, composes, and refines attack strategies through interaction and feedback, without updating model parameters. Specifically, SRTJ couples experience-driven attack generation with answer set programming (ASP)-based rule selection and constraint-aware composition, where iterative verifier feedback is leveraged to jointly refine successful strategies and analyze failure patterns. The resulting rule memory evolves in a hierarchical multi-level manner, explicitly organizing distilled attack knowledge into long-term, middle-term, and short-term rules, thereby capturing both stable transferable strategies and transient adaptive behaviors to effectively balance exploration and exploitation across attack attempts. Extensive experiments on mainstream jailbreak benchmark (HarmBench) demonstrate that SRTJ achieves strong and stable attack performance across different target LLMs, while exhibiting improved robustness and generalization compared to existing jailbreak methods. The code is available at https://github.com/TheSolkatt/SRTJ.

Safety alignment is critical for the responsible deployment of large language models (LLMs). As Mixture-of-Experts (MoE) architectures are increasingly adopted to scale model capacity, understanding their safety robustness becomes essential. Existing adversarial attacks, however, have notable limitations. Prompt-based jailbreaks rely on heuristic search and transfer poorly, model intervention methods require privileged access to internal representations, and optimization-based input attacks remain output-centric and are fundamentally limited to MoE models due to the non-differentiable routing mechanism. In this paper, we present RouteHijack, a routing-aware jailbreak for MoE LLMs. Our key insight is that safety behavior is concentrated in a small subset of experts, creating an opportunity to steer model behavior by influencing routing decisions through input optimization. Building on this observation, RouteHijack first performs response-driven expert localization to identify safety-critical and harmful experts by contrasting activations under safe refusals and harmful completions. It then constructs adversarial suffixes with a routing-aware objective that suppresses safety experts, promotes harmful experts, and prevents early-stage refusal during generation. At inference time, the optimized suffix is appended to a malicious prompt, requiring only input access. Across seven MoE LLMs, RouteHijack achieves a 69.3% average attack success rate (ASR), outperforming prior optimization-based attack by $3.2\times$. RouteHijack also transfers zero-shot across five sibling MoE variants, raising average ASR from 27.7% to 61.2%, and further generalizes to three MoE-based VLMs, increasing average ASR from 2.47% to 38.7%. These findings expose a fundamental vulnerability in sparse expert architectures and highlight the need for defenses beyond output-level alignment.

The visual modality of vision-language models (VLMs) is an underexplored attack surface for bypassing safety alignment. We introduce four jailbreak attacks exploiting the vision component: (1) encoding harmful instructions as visual symbol sequences with a decoding legend, (2) replacing harmful objects with benign substitutes (e.g., bomb -> banana) then prompting for harmful actions using the substitute term, (3) replacing harmful text in images (e.g., on book covers) with benign words while visual context preserves the original meaning, and (4) visual analogy puzzles whose solution requires inferring a prohibited concept. Evaluating across six frontier VLMs, our visual attacks bypass safety alignment and expose a cross-modality alignment gap: text-based safety training does not automatically generalize to harmful intent conveyed visually. For example, our visual cipher achieves 40.9% attack success on Claude-Haiku-4.5 versus 10.7% for an equivalent textual cipher. To further our insight into the attack mechanism, we present preliminary interpretability and mitigation results. These findings highlight that robust VLM alignment requires treating vision as a first-class target for safety post-training.

As language model safeguards become more robust, attackers are pushed toward developing increasingly complex jailbreaks. Prior work has found that this complexity imposes a "jailbreak tax" that degrades the target model's task performance. We show that this tax scales inversely with model capability and that the most advanced jailbreaks effectively yield no reduction in model capabilities. Evaluating 28 jailbreaks on five benchmarks across Claude models ranging in capability from Haiku 4.5 to Opus 4.6, we find Haiku 4.5 loses an average of 33.1% on benchmark performance when jailbroken, while Opus 4.6 at max thinking effort loses only 7.7%. We also observe that across all models, reasoning-heavy tasks display considerably more degradation than knowledge-recall tasks. Finally, Boundary Point Jailbreaking, currently the strongest jailbreak against deployed classifiers, achieves near-perfect classifier evasion with near-zero degradation across safeguarded models. We recommend that safety cases for frontier models should not rely on a meaningful capability degradation from jailbreaks.

Safety-aligned large language models rely on RLHF and instruction tuning to refuse harmful requests, yet the internal mechanisms implementing safety behavior remain poorly understood. We introduce the Attention Redistribution Attack (ARA), a white-box adversarial attack that identifies safety-critical attention heads and crafts nonsemantic adversarial tokens that redirect attention away from safety-relevant positions. Unlike prior jailbreak methods operating at the semantic or output-logit level, ARA targets the geometry of softmax attention on the probability simplex using Gumbel-softmax optimization over targeted heads. Across LLaMA-3-8B-Instruct, Mistral-7B-Instruct-v0.1, and Gemma-2-9B-it, ARA bypasses safety alignment with as few as 5 tokens and 500 optimization steps, achieving 36% ASR on Mistral-7B and 30% on LLaMA-3 against 200 HarmBench prompts, while Gemma-2 remains at 1%. Our principal mechanistic finding is a dissociation between ablation and redistribution: zeroing out the top-ranked safety heads produces at most 1 flip among 39 to 50 baseline refusals, while ARA targeting the corresponding safety-heavy layers flips 72/200 prompts on Mistral-7B and 60/200 on LLaMA-3. This suggests that safety is not localized in these heads as removable components, but emerges from the attention routing they perform. Removing a head allows compensation through the residual stream, while redirecting its attention propagates a corrupted signal downstream.

Safety trained large language models (LLMs) can often be induced to answer harmful requests through jailbreak prompts. Because we lack a robust understanding of why LLMs are susceptible to jailbreaks, future frontier models operating more autonomously in higher-stakes settings may similarly be vulnerable to such attacks. Prior work has studied jailbreak success by examining the model's intermediate representations, identifying directions in this space that causally encode concepts like harmfulness and refusal. Then, they globally explain all jailbreak attacks as attempting to reduce or strengthen these concepts (e.g., reduce harmfulness). However, different jailbreak strategies may succeed by strengthening or suppressing different intermediate concepts, and the same jailbreak strategy may not work for different harmful request categories (e.g., violence vs. cyberattack); thus, we seek to give a local explanation -- i.e., why did this specific jailbreak succeed? To address this gap, we introduce LOCA, a method that gives Local, CAusal explanations of jailbreak success by identifying a minimal set of interpretable, intermediate representation changes that causally induce model refusal on an otherwise successful jailbreak request. We evaluate LOCA on harmful original-jailbreak pairs from a large jailbreak benchmark across Gemma and Llama chat models, comparing against prior methods adapted to this setting. LOCA can successfully induce refusal by making, on average, six interpretable changes; prior work routinely fails to achieve refusal even after 20 changes. LOCA is a step toward mechanistic, local explanations of jailbreak success in LLMs. Code to be released.

Decompositional jailbreaks pose a critical threat to large language models (LLMs) by allowing adversaries to fragment a malicious objective into a sequence of individually benign queries that collectively reconstruct prohibited content. In real-world deployments, LLMs face a continuous, untraceable stream of fully anonymized and arbitrarily interleaved requests, infiltrated by covertly distributed adversarial queries. Under this rigorous threat model, state-of-the-art defensive strategies exhibit fundamental limitations. In the absence of trustworthy user metadata, they are incapable of tracking global historical contexts, while their deployment of generative models for real-time monitoring introduces computationally prohibitive overhead. To address this, we present TwinGate, a stateful dual-encoder defense framework. TwinGate employs Asymmetric Contrastive Learning (ACL) to cluster semantically disparate but intent-matched malicious fragments in a shared latent space, while a parallel frozen encoder suppresses false positives arising from benign topical overlap. Each request requires only a single lightweight forward pass, enabling the defense to execute in parallel with the target model's prefill phase at negligible latency overhead. To evaluate our approach and advance future research, we construct a comprehensive dataset of over 3.62 million instructions spanning 8,600 distinct malicious intents. Evaluated on this large-scale corpus under a strictly causal protocol, TwinGate achieves high malicious intent recall at a remarkably low false positive rate while remaining highly robust against adaptive attacks. Furthermore, our proposal substantially outperforms stateful and stateless baselines, delivering superior throughput and reduced latency.

Safety mechanisms for large language models (LLMs) remain predominantly English-centric, creating systematic vulnerabilities in multilingual deployment. Prior work shows that translating malicious prompts into other languages can substantially increase jailbreak success rates, exposing a structural cross-lingual security gap. We investigate whether such attacks can be mitigated through language-agnostic semantic similarity without retraining or language-specific adaptation. Our approach compares multilingual query embeddings against a fixed English codebook of jailbreak prompts, operating as a training-free external guardrail for black-box LLMs. We conduct a systematic evaluation across four languages, two translation pipelines, four safety benchmarks, three embedding models, and three target LLMs (Qwen, Llama, GPT-3.5). Our results reveal two distinct regimes of cross-lingual transfer. On curated benchmarks containing canonical jailbreak templates, semantic similarity generalizes reliably across languages, achieving near-perfect separability (AUC up to 0.99) and substantial reductions in absolute attack success rates under strict low-false-positive constraints. However, under distribution shift - on behaviorally diverse and heterogeneous unsafe benchmarks - separability degrades markedly (AUC $\approx$ 0.60-0.70), and recall in the security-critical low-FPR regime drops across all embedding models.

Existing white-box jailbreak attacks against aligned LLMs typically append discrete adversarial suffixes to the user prompt, which visibly alters the prompt and operates in a combinatorial token space. Prior work has avoided directly optimizing the embeddings of the original prompt tokens, presumably because perturbing them risks destroying the prompt's semantic content. We propose Prompt Embedding Optimization (PEO), a multi-round white-box jailbreak that directly optimizes the embeddings of the original prompt tokens without appending any adversarial tokens, and show that the concern is unfounded: the optimized embeddings remain close enough to their originals that the visible prompt string is preserved exactly after nearest-token projection, and quantitative analysis shows the model's responses stay on topic for the large majority of prompts. PEO combines continuous embedding-space optimization with structured continuation targets and an adaptive failure-focused schedule. Counterintuitively, later PEO rounds can benefit from heuristic composite response scaffolds that are not natural standalone templates, yet ASR-Judge shows that the resulting gains are not merely empty formatting or scaffold-only outputs. Across two standard harmful-behavior benchmarks and competing white-box attacks spanning discrete suffix search, appended adversarial embeddings, and search-based adversarial generation, PEO outperforms all of them in our experiments.

Large language models deployed at runtime can misbehave in ways that clean-data validation cannot anticipate: training-time backdoors lie dormant until triggered, jailbreaks subvert safety alignment, and prompt injections override the deployer's instructions. Existing runtime defenses address these threats one at a time and often assume a clean reference model, trigger knowledge, or editable weights, assumptions that rarely hold for opaque third-party artifacts. We introduce Layerwise Convergence Fingerprinting (LCF), a tuning-free runtime monitor that treats the inter-layer hidden-state trajectory as a health signal: LCF computes a diagonal Mahalanobis distance on every inter-layer difference, aggregates via Ledoit-Wolf shrinkage, and thresholds via leave-one-out calibration on 200 clean examples, with no reference model, trigger knowledge, or retraining. Evaluated on four architectures (Llama-3-8B, Qwen2.5-7B, Gemma-2-9B, Qwen2.5-14B) across backdoors, jailbreaks, and prompt injection (56 backdoor combinations, 3 jailbreak techniques, and BIPIA email + code-QA), LCF reduces mean backdoor attack success rate (ASR) below 1% on Qwen2.5-7B and Gemma-2 and to 1.3% on Qwen2.5-14B, detects 92-100% of DAN jailbreaks (62-100% for GCG and softer role-play), and flags 100% of text-payload injections across all eight (model, domain) cells, at 12-16% backdoor FPR and <0.1% inference overhead. A single aggregation score covers all three threat families without threat-specific tuning, positioning LCF as a general-purpose runtime safety layer for cloud-served and on-device LLMs.

Large (vision-)language models exhibit remarkable capability but remain highly susceptible to jailbreaking. Existing safety training approaches aim to have the model learn a refusal boundary between safe and unsafe, based on the user's intent. It has been found that this binary training regime often leads to brittleness, since the user intent cannot reliably be evaluated, especially if the attacker obfuscates their intent, and also makes the system seem unhelpful. In response, frontier models, such as GPT-5, have shifted from refusal-based safeguards to safe completion, that aims to maximize helpfulness while obeying safety constraints. However, safe completion could be exploited when a user pretends their intention is benign. Specifically, this intent inversion would be effective in multi-turn conversation, where the attacker has multiple opportunities to reinforce their deceptively benign intent. In this work, we introduce a novel multi-turn jailbreaking method that exploits this vulnerability. Our approach gradually builds conversational trust by simulating benign-seeming intentions and by exploiting the consistency property of the model, ultimately guiding the target model toward harmful, detailed outputs. Most crucially, our approach also uncovered an additional class of model vulnerability that we call para-jailbreaking that has been unnoticed up to now. Para-jailbreaking describes the situation where the model may not reveal harmful direct reply to the attack query, however the information that it reveals is nevertheless harmful. Our contributions are threefold. First, it achieves high success rates against frontier models including GPT-5-thinking and Claude-Sonnet-4.5. Second, our approach revealed and addressed para-jailbreaking harmful output. Third, experiments on multimodal VLM models showed that our approach outperformed state-of-the-art models.

The deployment of Large Language Models (LLMs) as assistants in electric grid operations promises to streamline compliance and decision-making but exposes new vulnerabilities to prompt-based adversarial attacks. This paper evaluates the risk of jailbreaking LLMs, i.e., circumventing safety alignments to produce outputs violating regulatory standards, assuming threats from authorized users, such as operators, who craft malicious prompts to elicit non-compliant guidance. Three state-of-the-art LLMs (OpenAI's GPT-4o mini, Google's Gemini 2.0 Flash-Lite, and Anthropic's Claude 3.5 Haiku) were tested against Baseline, BitBypass, and DeepInception jailbreaking methods across scenarios derived from nine NERC Reliability Standards (EOP, TOP, and CIP). In the initial broad experiment, the overall Attack Success Rate (ASR) was 33.1%, with DeepInception proving most effective at 63.17% ASR. Claude 3.5 Haiku exhibited complete resistance (0% ASR), while Gemini 2.0 Flash-Lite was most vulnerable (55.04% ASR) and GPT-4o mini moderately susceptible (44.34% ASR). A follow-up experiment refining malicious wording in Baseline and BitBypass attacks yielded a 30.6% ASR, confirming that subtle prompt adjustments can enhance simpler methods' efficacy.

Jailbreak attacks expose a persistent failure mode in safety-aligned LLMs: models can be pushed into harmful behavior, but the internal representations enabling this shift remain poorly localized. Recent mechanistic safety studies often explain such behavior through broad representational objects, including global refusal directions, activation steering vectors, and refusal-related SAE features. We instead ask whether jailbreak vulnerability can be traced to finer-grained, prompt-conditioned SAE feature subgroups. We introduce a token-driven mechanistic pipeline that decomposes the residual stream of Gemma-2-2B into Sparse Autoencoder (SAE) features and identifies feature subgroups associated with unsafe behavior. Using single-category unsafe examples from BeaverTails to reduce cross-category interference, we extract harmful concepts from adversarial responses and align them with concept-relevant prompt tokens through subspace similarity. We then apply three feature-grouping strategies: cluster-based, hierarchical-linkage, and single-token-driven, to identify SAE feature subgroups across all 26 layers. Finally, we amplify the top features in each subgroup and evaluate the resulting generations with a standardized harmfulness judge. Single-token-driven grouping achieves harmfulness comparable to full cluster-based grouping, showing that individual harmful prompt tokens are sufficient to localize vulnerability-relevant SAE feature subgroups without relying on broader cluster-level aggregation. These subgroups appear across early and mid-to-late layers, with stronger concentration in mid-to-late layers, where targeted steering exposes specific model vulnerabilities. Overall, our results suggest that jailbreak susceptibility can be traced to sparse, token-localized SAE feature subgroups, complementing prior accounts based on broad adversarial, refusal, or steering directions.

Large Language Models (LLMs) have been widely explored in educational scenarios. We identify a critical vulnerability in current educational LLMs, pedagogical jailbreaks, where students use answer-inducing prompts to elicit solutions rather than scaffolded instructions. To enable systematic study, we unify and formalize safe, helpful, and pedagogical behaviors with a knowledge-mastery graph and introduce SHAPE, a benchmark of 9,087 student-question pairs for evaluating tutoring behavior under adversarial pressure. We propose a graph-augmented tutoring pipeline that infers prerequisite concepts from queries, identifies mastery gaps, and routes generation between instructing and problem-solving via explicit gating. Experiments across multiple LLMs show that our method yields significantly improved safety under two pedagogical jailbreak settings, while maintaining near-ceiling helpfulness under the same evaluation protocol. Our code and data are available at https://github.com/MAPS-research/SHaPE

As artificial intelligence (AI) systems are increasingly deployed across critical domains, their security vulnerabilities pose growing risks of high-profile exploits and consequential system failures. Yet systematic approaches to evaluating AI security remain underdeveloped. In this paper, we introduce AVISE (AI Vulnerability Identification and Security Evaluation), a modular open-source framework for identifying vulnerabilities in and evaluating the security of AI systems and models. As a demonstration of the framework, we extend the theory-of-mind-based multi-turn Red Queen attack into an Adversarial Language Model (ALM) augmented attack and develop an automated Security Evaluation Test (SET) for discovering jailbreak vulnerabilities in language models. The SET comprises 25 test cases and an Evaluation Language Model (ELM) that determines whether each test case was able to jailbreak the target model, achieving 92% accuracy, an F1-score of 0.91, and a Matthews correlation coefficient of 0.83. We evaluate nine recently released language models of diverse sizes with the SET and find that all are vulnerable to the augmented Red Queen attack to varying degrees. AVISE provides researchers and industry practitioners with an extensible foundation for developing and deploying automated SETs, offering a concrete step toward more rigorous and reproducible AI security evaluation.

Effective safety auditing of large language models (LLMs) demands tools that go beyond black-box probing and systematically uncover vulnerabilities rooted in model internals. We present a comprehensive, interpretability-driven jailbreaking audit of eight SOTA open-source LLMs: Llama-3.1-8B, Llama-3.3-70B-4bt, GPT-oss- 20B, GPT-oss-120B, Qwen3-0.6B, Qwen3-32B, Phi4-3.8B, and Phi4-14B. Leveraging interpretability-based approaches -- Universal Steering (US) and Representation Engineering (RepE) -- we introduce an adaptive two-stage grid search algorithm to identify optimal activation-steering coefficients for unsafe behavioral concepts. Our evaluation, conducted on a curated set of harmful queries and a standardized LLM-based judging protocol, reveals stark contrasts in model robustness. The Llama-3 models are highly vulnerable, with up to 91% (US) and 83% (RepE) jailbroken responses on Llama-3.3-70B-4bt, while GPT-oss-120B remains robust to attacks via both interpretability approaches. Qwen and Phi models show mixed results, with the smaller Qwen3-0.6B and Phi4-3.8B mostly exhibiting lower jailbreaking rates, while their larger counterparts are more susceptible. Our results establish interpretability-based steering as a powerful tool for systematic safety audits, but also highlight its dual-use risks and the need for better internal defenses in LLM deployment.

Large language models (LLMs) are increasingly deployed in applications for global multilingual users, yet safety training remains concentrated in dominant languages and has not progressed in parallel with multilingual capability, creating exploitable gaps for jailbreak attacks. Current jailbreak defenses are largely developed and evaluated in dominant languages, and their effectiveness is limited by the scarcity of aligned multilingual supervision and representations dispersion caused by language variation. To address this issue, we propose MLJailDe, a multilingual jailbreak detection framework designed to improve both multilingual robustness and cross-lingual generalization. MLJailDe first introduces a multilingual back-translation data augmentation algorithm to construct a semantically consistent and functionally effective dataset spanning 11 languages, consisting of 2,232 benign and 1,239 jailbreak samples. On this basis, MLJailDe employs relative-distance constraints to reduce cross-lingual representation dispersion and encourage jailbreak prompts with similar intent to form consistent clusters across languages, while an imbalance-aware classification objective is further used to alleviate class imbalance and learn more reliable multilingual decision boundaries. Experimental results show that MLJailDe outperforms state-of-the-art baselines across multiple languages, achieving an F1 score of 98.5%, and obtains an average F1 score of 97.1% on unseen languages, demonstrating strong effectiveness and cross-lingual generalization.

As LLMs continue to shape real-world applications, automated jailbreak generation becomes essential to reveal safety weaknesses and guide model improvement. Existing automatic jailbreak generation methods have not yet fully considered two important aspects: adaptability to evolving safety-finetuned models, which affects their effectiveness on newer model versions, and diversity in generated prompts, which can cause narrow or repetitive attack patterns. To address these issues, we propose EvoJail, an instruction-fusion-driven evolutionary jailbreak generation framework that formalizes jailbreak prompt generation as a multi-objective black-box optimization problem and leverages the principles of evolutionary algorithms to search for jailbreak prompts that can adapt across different model versions and exhibit diverse attack patterns. Specifically, EvoJail integrates jailbreak prompt generation into an iterative evolutionary loop, where at each iteration candidate prompts are evaluated directly against the target model and then selected and varied based on the target model's responses, enabling the generation process to continuously adapt to model updates. To enhance diversity, EvoJail introduces field-aware instruction fusion to construct diverse starting points and incorporates diversity-aware objectives into the evolutionary fitness function, guiding the search toward prompts with richer semantic variation, while further designing multi-level LLM-based mutation operators that modify prompt structures at different granularities to promote structural diversity throughout the evolutionary process. Results demonstrate that EvoJail has stronger adaptability and can achieve over $93\%$ attack success rate and more than $5.6\%$ improvement in diversity metrics over state-of-the-art methods.

Large language models (LLMs) are increasingly used as co-authors in collaborative writing, where users begin with rough drafts and rely on LLMs to complete, revise, and refine their content. However, this capability poses a serious safety risk: malicious users could jailbreak the models-filling incomplete drafts with dangerous content-to force them into generating harmful outputs. In this paper, we identify the vulnerability of current LLMs to such draft-based co-authoring jailbreak attacks and introduce HarDBench, a systematic benchmark designed to evaluate the robustness of LLMs against this emerging threat. HarDBench spans a range of high-risk domains-including Explosives, Drugs, Weapons, and Cyberattacks-and features prompts with realistic structure and domain-specific cues to assess the model susceptibility to harmful completions. To mitigate this risk, we introduce a safety-utility balanced alignment approach based on preference optimization, training models to refuse harmful completions while remaining helpful on benign drafts. Experimental results show that existing LLMs are highly vulnerable in co-authoring contexts and our alignment method significantly reduces harmful outputs without degrading performance on co-authoring capabilities. This presents a new paradigm for evaluating and aligning LLMs in human-LLM collaborative writing settings. Our new benchmark and dataset are available on our project page at https://github.com/untae0122/HarDBench

While Large Language Models (LLMs) are widely used, they remain susceptible to jailbreak prompts that can elicit harmful or inappropriate responses. This paper introduces STAR-Teaming, a novel black-box framework for automated red teaming that effectively generates such prompts. STAR-Teaming integrates a Multi-Agent System (MAS) with a Strategy-Response Multiplex Network and employs network-driven optimization to sample effective attack strategies. This network-based approach recasts the intractable high-dimensional embedding space into a tractable structure, yielding two key advantages: it enhances the interpretability of the LLM's strategic vulnerabilities, and it streamlines the search for effective strategies by organizing the search space into semantic communities, thereby preventing redundant exploration. Empirical results demonstrate that STAR-Teaming significantly surpasses existing methods, achieving a higher attack success rate (ASR) at a lower computational cost. Extensive experiments validate the effectiveness and explainability of the Multiplex Network. The code is available at https://github.com/selectstar-ai/STAR-Teaming-paper.

Detecting jailbreak behaviour in large language models remains challenging, particularly when strongly aligned models produce harmful outputs only rarely. In this work, we present an empirical study of output based jailbreak detection under realistic conditions using the JailbreakBench Behaviors dataset and multiple generator models with varying alignment strengths. We evaluate both a lexical TF-IDF detector and a generation inconsistency based detector across different sampling budgets. Our results show that single output evaluation systematically underestimates jailbreak vulnerability, as increasing the number of sampled generations reveals additional harmful behaviour. The most significant improvements occur when moving from a single generation to moderate sampling, while larger sampling budgets yield diminishing returns. Cross generator experiments demonstrate that detection signals partially generalise across models, with stronger transfer observed within related model families. A category level analysis further reveals that lexical detectors capture a mixture of behavioural signals and topic specific cues, rather than purely harmful behaviour. Overall, our findings suggest that moderate multi sample auditing provides a more reliable and practical approach for estimating model vulnerability and improving jailbreak detection in large language models. Code will be released.

Large Language Models (LLMs) remain vulnerable to optimization-based jailbreak attacks that exploit internal gradient structure. While Sparse Autoencoders (SAEs) are widely used for interpretability, their robustness implications remain underexplored. We present a study of integrating pretrained SAEs into transformer residual streams at inference time, without modifying model weights or blocking gradients. Across four model families (Gemma, LLaMA, Mistral, Qwen) and two strong white-box attacks (GCG, BEAST) plus three black-box benchmarks, SAE-augmented models achieve up to a 5x reduction in jailbreak success rate relative to the undefended baseline and reduce cross-model attack transferability. Parametric ablations reveal (i) a monotonic dose-response relationship between L0 sparsity and attack success rate, and (ii) a layer-dependent defense-utility tradeoff, where intermediate layers balance robustness and clean performance. These findings are consistent with a representational bottleneck hypothesis: sparse projection reshapes the optimization geometry exploited by jailbreak attacks.

Open-weight language models can be rendered unsafe through several distinct interventions, but the resulting models may differ substantially in capabilities, behavioral profile, and internal failure mode. We study behavioral and mechanistic properties of jailbroken models across three unsafe routes: harmful supervised fine-tuning (SFT), harmful reinforcement learning with verifiable rewards (RLVR), and refusal-suppressing abliteration. All three routes achieve near-ceiling harmful compliance, but they diverge once we move beyond direct harmfulness. RLVR-jailbroken models show minimal degradation and preserve explicit harm recognition in a structured self-audit: they are able to identify harmful prompts and describe how a safe LLM should respond, yet they comply with the harmful request. With RLVR, harmful behavior is strongly suppressed by a reflective safety scaffold: when a harmful prompt is prepended with an instruction to reflect on safety standards, harmful behavior drops close to the baseline. Category-specific RLVR jailbreaks generalize broadly across harmfulness domains. Models jailbroken with SFT show the largest collapse in explicit safety judgments, the highest behavioral drift, and a substantial capability loss on standard benchmarks. Abliteration is family-dependent in both self-audit and response to a reflective safety scaffold. Mechanistic and repair analyses further separate the routes: abliteration is consistent with localized refusal-feature deletion, RLVR with preserved safety geometry but retargeted policy behavior, and SFT with broader distributed drift. Targeted repair partially recovers RLVR-jailbroken models, but has little effect on SFT-jailbroken models. Together, these results show that jailbreaks can produce vastly different properties despite similar harmfulness, with models jailbroken via RLVR showing remarkable similarity to the base model.

Pluralism alignment with AI has the sophisticated and necessary goal of creating AI that can coexist with and serve morally multifaceted humanity. Research towards pluralism alignment has many efforts in enhancing the learning of large language models (LLMs) to accomplish pluralism. Although this is essential, the robustness of LLMs to produce moral content over pluralistic values is still under exploration.Inspired by the astonishing persuasion abilities via jailbreak prompts, we propose to leverage jailbreak attacks to study LLMs' internal pluralistic values. In detail, we develop a morality dataset with 10.3K instances in two categories: Value Ambiguity and Value Conflict. We further formalize four adversarial attacks with the constructed dataset, to manipulate LLMs' judgment over the morality questions. We evaluate both the large language models and guardrail models which are typically used in generative systems with flexible user input. Our experiment results show that there is a critical vulnerability of LLMs and guardrail models to these subtle and sophisticated moral-aware attacks.

Multi-turn jailbreak attacks progressively erode LLM safety alignment across seemingly innocuous conversation turns, achieving success rates exceeding 90% against state-of-the-art models. Existing alignment-based and guardrail methods suffer from three key limitations: they require costly weight modification, evaluate each turn independently without modeling cumulative safety erosion, and detect attacks only after harmful content has been generated. To address these limitations, we first formulate the proactive early jailbreak detection problem with a new metric, detection lead, that measures how early an attack can be detected before the LLM complies. We then propose SAFEDREAM, a lightweight world-model-based framework that operates as an external module without modifying the LLM's weights. SAFEDREAM introduces three components: (1) a safety state world model that encodes LLM hidden states into a compact safety representation and predicts how it evolves across turns, (2) CUSUM detection that accumulates weak per-turn risk signals into reliable evidence, and (3) contrastive imagination that simultaneously rolls out attack and benign futures in latent space to issue early alarms before jailbreaks occur. On three multi-turn jailbreak benchmarks (XGuard-Train, SafeDialBench, SafeMTData) against 8 baselines, SAFEDREAM achieves the best detection timeliness across all benchmarks (1.06-1.20 turns before compliance) while maintaining competitive false positive rates and outperforming baselines in detection quality.

Text-to-image (T2I) models are increasingly optimized for following user instructions faithfully. However, we find that this capability introduces a safety vulnerability we call Multi-Concept Compositional Unsafety (MCCU). MCCU occurs when multiple individually safe concepts, if combined in a single generation request, lead to harmful or sensitive visual outputs. Unlike prior jailbreak settings, MCCU does not rely on adversarial prompts, model access, or explicitly disallowed content. Instead, the risk emerges from how the model composes multiple safe visual concepts into a single scene. To systematically measure this threat, we build TwoHamsters, a large-scale evaluation framework consisting of 20k prompts, 51 curated concept pairs, and six risk categories. We evaluate 13 T2I models under a black-box setting. Our results show a clear conflict between instruction-following and safety: models that follow prompts more faithfully tend to produce more MCCU failures. For example, FLUX.1 achieves a 99.35% Unsafe Alignment Rate while only reaching a 1.57% MCCU Defense Rate. We further evaluate three representative defenses, including safety filtering, MCCU-specific detector fine-tuning, and concept erasure, all of which fail against unseen concept combinations. Our findings suggest that compositional reasoning in T2I models creates an attack surface that is not captured by existing safety mechanisms. We anticipate the release of TwoHamsters will catalyze community development of advanced generative defense mechanisms.

Large Reasoning Models (LRMs) have demonstrated strong capabilities in generating step-by-step reasoning chains alongside final answers, enabling their deployment in high-stakes domains such as healthcare and education. While prior jailbreak attack studies have focused on the safety of final answers, little attention has been given to the safety of the reasoning process. In this work, we identify a novel problem that injects harmful content into the reasoning steps while preserving unchanged answers. This type of attack presents two key challenges: 1) manipulating the input instructions may inadvertently alter the LRM's final answer, and 2) the diversity of input questions makes it difficult to consistently bypass the LRM's safety alignment mechanisms and embed harmful content into its reasoning process. To address these challenges, we propose the Psychology-based Reasoning-targeted Jailbreak Attack (PRJA) Framework, which integrates a Semantic-based Trigger Selection module and a Psychology-based Instruction Generation module. Specifically, the proposed PRJA automatically selects manipulative reasoning triggers via semantic analysis and leverages psychological theories of obedience to authority and moral disengagement to generate adaptive instructions for enhancing the LRM's compliance with harmful content generation. Extensive experiments on five question-answering datasets demonstrate that PRJA achieves an average attack success rate of 83.6% against several commercial LRMs, including DeepSeek R1, Qwen2.5-Max, and OpenAI o4-mini.

As Spoken Language Models (SLMs) integrate speech and text modalities, they inherit the safety vulnerabilities of their LLM backbone while introducing an expanded attack surface. SLMs have been previously shown to be susceptible to jailbreaking, where adversarial prompts induce harmful responses. Yet existing attacks largely remain unimodal, optimizing either text or audio in isolation. We explore gradient-based multimodal jailbreaks by introducing JAMA (Joint Audio-text Multimodal Attack), a joint multimodal optimization framework combining Greedy Coordinate Gradient (GCG) for text and Projected Gradient Descent (PGD) for audio, to simultaneously perturb both modalities. Evaluations across four state-of-the-art SLMs and four audio types demonstrate that JAMA surpasses unimodal jailbreak rate by 1.5x to 20x. We analyze the operational dynamics of this joint attack and show that a sequential approximation method makes it 4x to 6x faster. Our findings suggest that unimodal safety is insufficient for robust SLMs. The code and data are available at https://repos.lsv.uni-saarland.de/akrishnan/multimodal-jailbreak-slm.

Jailbreak attacks remain a persistent threat to large language model safety. We propose Context-Conditioned Delta Steering (CC-Delta), an SAE-based defense that identifies jailbreak-relevant sparse features by comparing token-level representations of the same harmful request with and without jailbreak context. Using paired harmful/jailbreak prompts, CC-Delta selects features via statistical testing and applies inference-time mean-shift steering in SAE latent space. Across four aligned instruction-tuned models and twelve jailbreak attacks, CC-Delta achieves comparable or better safety-utility tradeoffs than baseline defenses operating in dense latent space. In particular, our method clearly outperforms dense mean-shift steering on all four models, and particularly against out-of-distribution attacks, showing that steering in sparse SAE feature space offers advantages over steering in dense activation space for jailbreak mitigation. Our results suggest off-the-shelf SAEs trained for interpretability can be repurposed as practical jailbreak defenses without task-specific training.

Recent advances in large image editing models have shifted the paradigm from text-driven instructions to vision-prompt editing, where user intent is inferred directly from visual inputs such as marks, arrows, and visual-text prompts. While this paradigm greatly expands usability, it also introduces a critical and underexplored safety risk: the attack surface itself becomes visual. In this work, we propose Vision-Centric Jailbreak Attack (VJA), the first visual-to-visual jailbreak attack that conveys malicious instructions purely through visual inputs. To systematically study this emerging threat, we introduce IESBench, a safety-oriented benchmark for image editing models. Extensive experiments on IESBench demonstrate that VJA effectively compromises state-of-the-art commercial models, achieving attack success rates of up to 80.9% on Nano Banana Pro and 70.1% on GPT-Image-1.5. To mitigate this vulnerability, we propose a training-free defense based on introspective multimodal reasoning, which substantially improves the safety of poorly aligned models to a level comparable with commercial systems, without auxiliary guard models and with negligible computational overhead. Our findings expose new vulnerabilities, provide both a benchmark and practical defense to advance safe and trustworthy modern image editing systems. Warning: This paper contains offensive images created by large image editing models.

Vision-Language Models (VLMs) are now a core part of modern AI. Recent work proposed several visual jailbreak attacks using single/ holistic images. However, contemporary VLMs demonstrate strong robustness against such attacks due to extensive safety alignment through preference optimization, e.g., reinforcement learning from human feedback (RLHF). In this work, we identify a new vulnerability: while VLM pretraining and instruction tuning generalize well to split-image inputs, safety alignment is typically performed only on holistic images and does not account for harmful semantics distributed across multiple image fragments. Consequently, VLMs often fail to detect and reject harmful split-image inputs, in which unsafe cues emerge only upon combining images. We introduce novel split-image visual jailbreak attacks (\textbf{SIVA}) that exploit this misalignment. Unlike prior optimization-based attacks, which exhibit poor black-box transferability due to architectural and prior mismatches across models, our attacks evolve in progressive phases from naive splitting to an adaptive white-box attack, culminating in a black-box transfer attack. Our strongest strategy leverages a novel adversarial knowledge distillation \textbf{(Adv-KD)} algorithm to substantially improve cross-model transferability. Evaluations on four state-of-the-art modern VLMs and three jailbreak datasets demonstrate that our strongest attack achieves up to 44% higher transfer success than existing baselines. Lastly, we propose efficient ways to address this critical vulnerability in the current VLM safety alignment.

Large Language Models (LLMs) are increasingly used as interfaces to information, code, and real-world services, making prompt-level security failures a practical concern. Although jailbreak attacks, defenses, datasets, and automated judgers have advanced rapidly, evaluation remains fragmented across threat models, access assumptions, cost budgets, datasets, and success criteria. This makes reported attack success rates and defense gains hard to compare. This SoK systematizes LLM prompt security across concepts, data, tooling, and measurement. We propose linked taxonomies for jailbreak attacks, defenses, and model vulnerabilities, while separating technical mechanisms from attacker and defender capabilities. We also formalize threat, access, and cost assumptions as explicit evaluation metadata. To support reproducible evaluation, we release JailbreakDB, PromptSecurity-Eval, and PromptSecurity, a modular platform that represents each experiment as a tuple of model, attack, defense, dataset, and judger. Using matched evaluations across models, attacks, defenses, and judgers, we show that access regime, native harmful-query behavior, attack cost, defense backfire, taxonomy subcategory, and judger choice all materially affect security conclusions. Together, these artifacts support reproducible, cost-aware, and taxonomy-grounded evaluation of LLM prompt security. Leaderboard: https://datasec-lab.github.io/PromptSecurityLeaderboard/. Dataset: https://huggingface.co/datasets/youbin2014/JailbreakDB. GitHub: https://github.com/datasec-lab/PromptSecurity.

Large Language Model (LLM) agents are powering a growing share of interactive web applications, yet remain vulnerable to misuse and harm. Prior jailbreak research has largely focused on single-turn prompts, whereas real harassment often unfolds over multi-turn interactions. In this work, we present the Online Harassment Agentic Benchmark consisting of: (i) a synthetic multi-turn harassment conversation dataset, (ii) a multi-agent (e.g., harasser, victim) simulation informed by repeated game theory, (iii) three jailbreak methods attacking agents across memory, planning, and fine-tuning, and (iv) a mixed-methods evaluation framework. We utilize two prominent LLMs, LLaMA-3.1-8B-Instruct (open-source) and Gemini-2.0-flash (closed-source). Our results show that jailbreak tuning makes harassment nearly guaranteed with an attack success rate of 95.78--96.89% vs. 57.25--64.19% without tuning in Llama, and 99.33% vs. 98.46% without tuning in Gemini, while sharply reducing refusal rate to 1-2% in both models. The most prevalent toxic behaviors are Insult with 84.9--87.8% vs. 44.2--50.8% without tuning, and Flaming with 81.2--85.1% vs. 31.5--38.8% without tuning, indicating weaker guardrails compared to sensitive categories such as sexual or racial harassment. Qualitative evaluation further reveals that attacked agents reproduce human-like aggression profiles, such as Machiavellian/psychopathic patterns under planning, and narcissistic tendencies with memory. Counterintuitively, closed-source and open-source models exhibit distinct escalation trajectories across turns, with closed-source models showing significant vulnerability. Overall, our findings show that multi-turn and theory-grounded attacks not only succeed at high rates but also mimic human-like harassment dynamics, motivating the development of robust safety guardrails to ultimately keep online platforms safe and responsible.

Even modern AI models often remain vulnerable to multimodal queries in which harmful intent is embedded in images. A widely used approach for safety alignment is training with extensive multimodal safety datasets, but the costs of data curation and training are often prohibitive. To mitigate these costs, inference-time alignment has recently been explored, but they often lack generalizability across diverse multimodal jailbreaks and still incur notable overhead due to extra forward passes for response refinement or heavy pre-deployment calibration procedures. Here, we identify insufficient visual attention to safety-critical image regions as one of the key causes of multimodal safety failures. Building on this insight, we propose Multimodal Risk-Adaptive Steering (MoRAS), which enhances safety-critical visual attention via concise visual contexts for accurate multimodal risk assessment. This risk signal enables risk-adaptive steering for direct refusals, reducing inference overhead while remaining generalizable across diverse multimodal jailbreaks. Notably, MoRAS requires only a small calibration set to estimate multimodal risk, substantially reducing pre-deployment overhead. We conduct various empirical validations across multiple benchmarks and MLLM backbones, and observe that the proposed MoRAS consistently mitigates jailbreaks, preserves utility, and reduces computational overhead compared to state-of-the-art inference-time defenses.

Unlike regular tokens derived from existing text corpora, special tokens are artificially created to annotate structured conversations during the fine-tuning process of Large Language Models (LLMs). Serving as metadata of training data, these tokens play a crucial role in instructing LLMs to generate coherent and context-aware responses. We demonstrate that special tokens can be exploited to construct four attack primitives, with which malicious users can reliably bypass the internal safety alignment of online LLM services and circumvent state-of-the-art (SOTA) external content moderation systems simultaneously. Moreover, we found that addressing this threat is challenging, as aggressive defense mechanisms-such as input sanitization by removing special tokens entirely, as suggested in academia-are less effective than anticipated. This is because such defense can be evaded when the special tokens are replaced by regular ones with high semantic similarity within the tokenizer's embedding space. We systemically evaluated our method, named MetaBreak, on both lab environment and commercial LLM platforms. Our approach achieves jailbreak rates comparable to SOTA prompt-engineering-based solutions when no content moderation is deployed. However, when there is content moderation, MetaBreak outperforms SOTA solutions PAP and GPTFuzzer by 11.6% and 34.8%, respectively. Finally, since MetaBreak employs a fundamentally different strategy from prompt engineering, the two approaches can work synergistically. Notably, empowering MetaBreak on PAP and GPTFuzzer boosts jailbreak rates by 24.3% and 20.2%, respectively.

Existing gradient-based jailbreak attacks on Large Language Models (LLMs) typically optimize adversarial suffixes to align the LLM output with predefined target responses. However, restricting the objective as inducing fixed targets inherently constrains the adversarial search space, limiting the overall attack efficacy. Furthermore, existing methods typically require numerous optimization iterations to fulfill the large gap between the fixed target and the original LLM output, resulting in low attack efficiency. To overcome these limitations, we propose NonTextual Target Attack (NTA), the first gradient-based attack that relies on a non-textual constrained objective to maximize the unsafety probability of the LLM output, without enforcing any response patterns. For tractable optimization, we further decompose this objective into two constrained sub-objectives, which can be approximated by two differentiable unconstrained losses, to iteratively optimize the response and the adversarial prompt in the neighborhood of the original prompt, with a theoretical analysis to validate the decomposition. In contrast to existing attacks, NTA first realizes gradient-based prompt optimization on a non-textual target and significantly expands the attack space, enabling more flexible and efficient exploration of LLM vulnerabilities. Extensive evaluations show that \textsc{NTA} achieves an average attack success rate of 96.8% against recent safety-aligned LLMs with only 100 optimization iterations on AdvBench, outperforming state-of-the-art gradient-based attacks by over 40%.

Code-capable large language model (LLM) agents are embedded in software engineering workflows where they can read, write, and execute code, raising "jailbreak" stakes beyond text-only settings. Prior evaluations emphasize refusal or harmful-text detection, leaving open whether agents compile and run malicious programs. We present JAWS-Bench (Jailbreaks Across WorkSpaces), a benchmark spanning three escalating workspace regimes mirroring attacker capability: empty (JAWS-0), single-file (JAWS-1), and multi-file (JAWS-M). We pair it with a hierarchical, executable-aware Judge Framework that tests (i) compliance, (ii) attack success, (iii) syntactic correctness, and (iv) runtime executability to measure deployable harm. Across seven LLM backends from five families, prompt-only attacks in JAWS-0 achieve 61% compliance; 58% are harmful, 52% parse, and 27% run end-to-end. In JAWS-1, compliance reaches ~100% for stronger models with a mean ASR (Attack Success Rate) ~71%; JAWS-M raises mean ASR to ~75%, with 32% runnable attack code. Wrapping an LLM in an agent increases ASR by 1.6$\times$, by overturning initial refusals during planning and tool use. Additional evaluations with SWE-Agent and OpenAI Codex exhibit similar trends, indicating that JAWS-Bench can be reused across multiple agent frameworks. Category analyses identify which attack classes are most vulnerable and deployable, motivating execution-aware defenses and refusal-preserving agent designs.