From Shallow to Deep: Pinning Semantic Intent via Causal GRPO

Large language models (LLMs) remain vulnerable to jailbreak attacks that conceal harmful intent within complex adversarial prompts. Existing defenses primarily rely on input perturbation or harmful-output suppression, but they rarely model where malicious intent resides, resulting in brittle protection and excessive over-refusal. We propose SENTINEL, a plug-and-play, generation-time jailbreak defense that reframes mitigation as an intent extraction problem. Our key insight is that instruction-tuned LLMs exhibit strong input--output semantic consistency: regardless of jailbreak complexity, generated outputs tend to align with the attacker's true intent. SENTINEL exploits this property by matching semantically aligned input--output regions to extract intention-revealing subsequences, scores these subsequences using refusal-direction projections to estimate harmfulness, and halts generation when necessary. Experiments on HarmBench across multiple LLMs show that SENTINEL reduces jailbreak success rates to close to 5% while maintaining low over-refusal. We further demonstrate robustness to adaptive attacks and provide a mechanistic interpretation: SENTINEL re-distributes jailbreak features from alignment blind spots to aligned regions.

Currently, open-sourced large language models (OSLLMs) have demonstrated remarkable generative performance. However, as their structure and weights are made public, they are exposed to jailbreak attacks even after alignment. Existing attacks operate primarily at shallow levels, such as the prompt or embedding level, and often fail to expose vulnerabilities rooted in deeper model components, which creates a false sense of security for successful defense. In this paper, we propose Safety Attention Head Attack (SAHA), an attention-head-level jailbreak framework that explores the vulnerability in deeper but insufficiently aligned attention heads. SAHA contains two novel designs. Firstly, we reveal that deeper attention layers introduce more vulnerability against jailbreak attacks. Based on this finding, SAHA introduces Ablation-Impact Ranking head selection strategy to effectively locate the most vital layer for unsafe output. Secondly, we introduce a boundary-aware perturbation method, i.e. Layer-Wise Perturbation, to probe the generation of unsafe content with minimal perturbation to the attention. This constrained perturbation guarantees higher semantic relevance with the target intent while ensuring evasion. Extensive experiments show the superiority of our method: SAHA improves ASR by 14% over SOTA baselines, revealing the vulnerability of the attack surface on the attention head. Our code is available at https://anonymous.4open.science/r/SAHA.

Recent advances in alignment techniques such as Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO) have improved the safety of large language models (LLMs). However, these LLMs remain vulnerable to jailbreak attacks that disguise harmful intent through indirect or deceptive phrasing. Using causal intervention, we empirically demonstrate that this vulnerability stems from shallow alignment mechanisms that lack deep reasoning, often rejecting harmful prompts without truly understanding why they are harmful. To mitigate this vulnerability, we propose enhancing alignment through reasoning-aware post-training. We construct and release a novel Chain-of-Thought (CoT) fine-tuning dataset that includes both utility-oriented and safety-critical prompts with step-by-step rationales. Fine-tuning on this dataset encourages models to produce principled refusals grounded in reasoning, outperforming standard SFT baselines. Furthermore, inspired by failure patterns in CoT fine-tuning, we introduce Alignment-Weighted DPO, which targets the most problematic parts of an output by assigning different preference weights to the reasoning and final-answer segments. This produces finer-grained, targeted updates than vanilla DPO and improves robustness to diverse jailbreak strategies. Extensive experiments across multiple safety and utility benchmarks show that our method consistently improves alignment robustness while maintaining overall model utility.