This work identifies batch prompting as a distinct safety failure mode, not reducible to known vulnerabilities such as in-context learning or long-context effects, and analyzes its causes from two complementary perspectives: alignment signal weakening and refusal signal dilution.
Abstract
Batch prompting is a practical inference strategy for large language models, but its safety implications remain underexplored. We show that the success of batch prompting for utility does not extend to safety: a harmful question that is reliably refused in isolation can elicit a harmful response when embedded in a batch of benign questions. We identify this as a distinct safety failure mode, not reducible to known vulnerabilities such as in-context learning or long-context effects, and analyze its causes from two complementary perspectives: alignment signal weakening and refusal signal dilution. Across widely used open-source and frontier commercial models, batch prompting consistently achieves high attack success rates as a simple black-box attack. We further show that batch-aware preference optimization effectively mitigates the vulnerability. These findings highlight a blind spot in current safety alignment and point to batch-aware alignment as a necessary step toward robust deployment.
Experiments across three instruction-tuned models show that HiRoute achieves high safety rates across multiple safety benchmarks while preserving safe-response helpfulness, reducing over-refusal, and maintaining competitive performance on general-purpose tasks.
Fangzhou Chen, Shiji Zhao, Mengyan Wang et al.· 0 citations
LongGuard is presented, a framework that evaluates, mechanistically analyzes, and mitigates long-context guardrail failure, and proposes two training-free mitigations - Chunked Detection and Attention-Head Sharpening (AHS) - and a deployment protocol that selects configurations by context length and audit side.
Safety evaluation is critical for assessing whether aligned Large Language Models (LLMs) remain robust against jailbreak attacks. Existing automated testing methods, however, largely rely on response-level feedback: each candidate prompt typically requires generating a target-model response to evaluate its attack effectiveness. This process is expensive and, more importantly, provides only sparse guidance on strongly aligned models, where most candidates are rejected with the same failure outcome. This paper presents NeuronFuzz, a white-box fuzzing framework that exploits internal safety neurons as continuous execution feedback for LLM safety evaluation. A SafetyOracle converts safety-neuron activations into a continuous safety alarm score that serves as feedback for fuzzing and can be obtained during prefill, eliminating response generation from the fuzzing loop. To construct the SafetyOracle, NeuronFuzz uses template-invariant harmful and benign inputs and stability-aware selection to identify a compact set of safety neurons whose activations capture harmful-intent recognition. Moreover, since the safety alarm score is differentiable, NeuronFuzz uses its gradients to identify safety-sensitive template positions and a masked language model to generate fluent, context-compatible mutations while preserving original harmful payload and avoiding additional optimization variables. We evaluate NeuronFuzz across 21 text and multimodal models. Across five white-box source models, it achieves a 76-100% jailbreak discovery rate, outperforming baselines by up to 48 percentage points. Its optimized templates further transfer zero-shot to open-weight and six proprietary target models, achieving average ASR and top-5 ensemble ASR (EASR) of 69.6%/92.6% and 44.1%/60.0%, respectively.
Zhiyuan Xu, Muhammad Firhard Roslan, Joseph Gardiner et al.· 0 citations
This paper proposes SafeKeep, an inference-time safeguard that decouples safety judgment from tool execution: it assesses requests using flattened textual tool specifications while retaining the original schema-formatted specifications for execution.
Minghui Pan, Jiayuxuan Yang, Yuanyuan Yuan et al.· 0 citations
Model merging has become the default way to give an aligned language model new skills without retraining: a practitioner folds task vectors from math, code, or domain specialists into a safety-aligned base using task arithmetic, TIES, or DARE. This convenience is known to carry a safety cost, but almost all of that evidence rests on static refusal tests: fixed harmful prompts scored for compliance. We argue this is misleading. Because safety alignment is"shallow,"concentrated in the first few generated tokens, a merged model's static refusal can stay clean while a real adaptive attack still breaks it. We introduce SkillSafe-Bench, a controlled benchmark that scores skill-merged models on static refusal, adaptive jailbreak robustness, and capability retention under a conservative two-judge AND rule. Across six open-weight bases (five families, two scales), static safety does not predict robustness to attack: under a semantic template attack, safe-looking merges on the fragile bases (both Qwen scales and Gemma) are jailbroken 60-76% of the time while others (Llama, Phi-4) stay robust. We further show the static effect of merging is base-conditional, characterize same-recipe abliteration-style safety erosion through a data-free geometric signal (the overlap of a task vector with a safety subspace), and outline SubSafe-Merge, which projects this overlap away to remove that erosion at held capability. Adaptive evaluation is not optional for merged LLMs: the models that most need it look safe under static screening.
Large language models (LLMs) are safety-aligned before deployment to reduce harmful content generation. Yet neuron-level pruning attacks show that refusal can depend on a small set of removable units: disabling them can remove safety behavior while leaving much of the model usable. To address this problem, we introduce Mask2Shield (M2S), a masked-forward alignment method that trains a model under this functional pruning. The masked student must recover a safe refusal through the remaining computation, while a frozen, unmasked teacher supplies complete benign answers to limit capability drift. Across ten model configurations, M2S reduces successful recomputed pruning attacks from 80--279 to 1--44 out of 313 prompts while generally preserving four capability benchmarks. We also evaluate M2S with TwinBreak, which uses a different neuron-selection rule and iterative pruning procedure. Together, these results show that M2S makes targeted pruning less effective by reducing reliance on a small, removable safety-neuron set.
Jincheng Ying, Minghui Xu, Yinhao Xiao et al.· 0 citations