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.
Abstract
Safety guardrails serve as the last line of defense against harmful inputs and outputs of large language models (LLMs), yet they are trained and evaluated almost exclusively on short text. We present LongGuard, a framework that evaluates, mechanistically analyzes, and mitigates long-context guardrail failure. We formulate the task as Safety Needle-in-a-Haystack (SafetyNIAH) over a 0.25k-32k length grid; across 15 mainstream guardrails, unsafe recall drops monotonically by more than 50% on average, and a paired Benign-Fill vs. Needle-Repeat design attributes the failure to proportional dilution of the unsafe needle rather than to absolute length. A three-layer attention-logit-behavior analysis on six guardrails locates the mechanism: attention mass on the unsafe needle is diluted, the unsafe-over-safe logit margin is compressed in lockstep, and the detection decision collapses accordingly, with this attention->logit->behavior chain remaining consistent after partialling out length. We further isolate a sparse set of guard-specialized retrieval heads that exhibit partial specificity relative to their base models. Building on the analysis, we propose two training-free mitigations - Chunked Detection (CD) and Attention-Head Sharpening (AHS) - and a deployment protocol, Context-Aware Hyperparameter Routing (CAHR), that selects configurations by context length and audit side. Across five benchmarks spanning synthetic data, long-context attacks, and reasoning-model outputs, CAHR-CD and CAHR-AHS improve the six-guardrail average by 22% and 13%, respectively. Code and data are available online.
To mitigate the refusal-cue shortcut, sparse complementary masking is adapted as a lightweight post-hoc intervention that identifies and suppresses a small set of shortcut-associated attention heads and MLP neurons without retraining and achieves an approximately 79% relative reduction in response-initial detection failures induced by refusal cues, while preserving standard detection performance.
Yu Feng, Chun-Ting Zang, Chen Shen et al.· 0 citations
Large language models (LLMs) are increasingly deployed in safety-critical applications, yet jailbreak attacks can conceal harmful intent through role-playing, fictional scenarios, or seemingly benign motivations. Existing inference-time defenses may miss disguised attacks or excessively refuse legitimate requests. We propose SRD-GUARD, a parameter-free, black-box defense framework that exposes concealed intent through semantic rewriting and consensus-based risk assessment. Given an input prompt, SRD-GUARD generates five semantically related rewrites that preserve the underlying objective while removing unnecessary contextual packaging. The original prompt and rewrites are jointly evaluated by multiple independent LLM-based safety scorers on a continuous risk scale. A decision module combines absolute risk thresholds with relative risk changes between the original and rewritten prompts to adaptively intercept, preserve, or warn on requests. We evaluate SRD-GUARD against UNIATTACK, CIPHER, and DeepInception on Llama-3-8B-Uncensored and DeepSeek-V4-Flash using AdvBench and OR-Bench-Hard. SRD-GUARD achieves average DSRs of 91.44% and 100%, with ORRs of 8.00% and 12.00%, respectively. Compared with evaluated baselines, it provides a more favorable DSR--ORR trade-off. Ablation studies show that rewriting exposes concealed harmful intent, joint scoring improves robustness to individual evaluator behavior, and risk-adaptive decision making enables selective handling of ambiguous inputs. These results demonstrate that semantic intent exposure, consensus-based risk assessment, and relative-risk-aware routing provide an effective and selective approach to black-box jailbreak defense. The artifact is available at https://anonymous.4open.science/status/CICD-Guard-D648.
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.
Kihyun Kim, Heeseon Kim, Wonjun Lee et al.· 0 citations
Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks'CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.
Despite extensive alignment efforts, Large Language Models (LLMs) remain vulnerable to generating unsafe content under adversarial prompting, yet the internal mechanisms by which safety behaviors are implemented remain poorly understood. We study LLM safety from a mechanistic interpretability perspective and characterize a multi-stage *safety circuit* that organizes refusal behavior, consisting of (i) $\textbf{Harmful Detection Heads}$ that respond to harmful inputs, (ii) $\textbf{Safety Neurons}$ that mediate and stabilize safety signals in the residual stream, and (iii) $\textbf{Refusal Heads}$ that translate these signals into safe response generation. Using targeted attention-head and neuron-level interventions, we provide causal evidence consistent with this circuit organization, showing that suppressing upstream Harmful Detection Heads disrupts downstream refusal behavior and that safety neurons mediate this interaction. We validate that this decomposition recurs across multiple LLM architectures and adversarial attack settings, and use simple, architecture-preserving weight scaling as a mechanistic probe to test its functional relevance. Across six LLMs, circuit-guided scaling improves safety rates under attacks by 26.5%, while incurring only a 1.7% accuracy drop across four standard benchmarks. Overall, our results support a circuit-level interpretation of LLM safety and suggest that mechanistic abstractions can reveal stable and transferable patterns underlying aligned behavior.
Mixture-of-Experts (MoE) is a scaling architecture for large language models that activates only a small subset of expert modules per token, enabling massive parameter growth with nearly constant computation. Recent Hybrid MoE architecture adds \textit{shared experts} to capture consistently useful representations, further improving stability and generalization. MoE now powers many flagship open-source and commercial models, yet remains vulnerable to adversarial attacks. Specifically, sparse routing introduces a structural vulnerability: MoE safety hinges on which experts are activated, and adversaries can subvert this selection through jailbreak prompts, malicious fine-tuning, and weight-level pruning of safety-critical neurons. Existing defenses primarily focus on hardening the router, but an adversary may still manipulate or bypass the routing trajectory due to the routing process's nondeterministic nature, thereby collapsing the defense. To cope with this problem, we first identify theoretically and empirically that shared expert, an always-activated component containing a small proportion of safety-critical neurons, can overcome the uncertainty of sparsely activated routing path and serve as a router-independent anchor to enhance global safety alignment. Based on this insight, we propose SEAL, a training-time parameter-efficient defense that produces a plug-and-play adapter attached to shared expert, and SEAL++, a variant that adds an orthogonal constraint preserving pre-existing safety subspaces during training. We evaluate SEAL and SEAL++ across six attack scenarios that combine three adversarial inputs (harmful prompting, jailbreak, malicious fine-tuning) with and without neuron pruning. SEAL reduces attack success rate (ASR) by up to 60\%, at a capability cost of at most 1.4\% on a five-benchmark average. Additionally, SEAL can seamlessly integrate with router-level ......
Qingyu Meng, Yiwei Zha, Jiahuan Pei et al.· 0 citations
