Stylistic jailbreaks in LLaMA-3.1-8B suppress linear refusal direction (r=−0.65) while leaving harmful intent recoverable at 51.8%, and resist linear patching
Auto-published from arXiv:2611.18601 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.88, flags: [possible-zero-day])
Category
alignment
Model
LLaMA-3.1-8B; replicated on Gemma-2-9b-it
Surface
White-box model access (residual stream, attention heads; mechanistic interpretability experiments)
Setup
Paired dataset of Vanilla Harmful, Stylized Attack (successful bypasses), Vanilla Safe, and Stylized Safe prompts. A style-orthogonal refusal vector (v_pure_ref) is isolated on a held-out split following Arditi et al. (2024). Three probing experiments at Layer 15: (A) OOD linear probe trained only on vanilla prompts evaluated on stylized attacks; (B) cosine projection of stylized vs. vanilla harmful prompts onto v_pure_ref; (D) causal patching of v_pure_ref into the final-token residual stream across Layers 10–20 for verified bypasses. Full construction details in Appendix D.1–D.5; replication on Gemma-2-9b-it in Appendix D.7.
Reproduction anchor
code: https://github.com/Sar-Hal/StylicJailBreaking; full mechanistic procedure in Appendix D (sections D.1–D.7, including Table 11 for Gemma-2-9b-it replication)
Observed behavior
Experiment A: OOD probe detects harmful intent in stylized attacks at 51.8% with 0.0% FPR on stylized-safe controls—intent is not erased. Experiment B: Stylized attacks project significantly less onto v_pure_ref than vanilla harmful prompts (Mann-Whitney p=3.6×10^-9, r=−0.65 at Layer 15). Experiment D: Patching v_pure_ref across Layers 10–20 restores refusal in only 8% of verified bypasses, with no single layer higher; zeroing v_pure_ref on vanilla harmful prompts produces no change (all still refuse). Full replication on Gemma-2-9b-it yields p=2.6×10^-10, r=−0.69, and linear patching restores only 4.5%.
Expected behavior
If refusal were mediated by a single linear direction, patching v_pure_ref should restore refusal in the majority of bypasses. If style worked via attention hijacking, safety heads should show disproportionate payload-attention drops relative to non-safety control heads. Neither is observed.
Reproducibility
high
Threat model
Defenders building single-direction activation-steering or attention-head ablation countermeasures against jailbreaks will find them ineffective against stylistic attacks. This failure mode operates via a distributed non-linear mechanism outside the primary linear refusal subspace, meaning representation-engineering defenses deployed in production inference pipelines (e.g., refusal-direction clamping) provide no protection against this attack class.
Novelty
Mechanistically rules out both semantic-camouflage and attention-hijacking as explanations for stylistic jailbreaks, demonstrating instead a distributed non-linear disruption of refusal that replicates across two distinct architectures—establishing a new mechanistic category of alignment failure that requires preference-training-based rather than linear-steering defenses.
Source
- arXiv: 2611.18601
- PDF: https://arxiv.org/pdf/2611.18601v1
- Categories: cs.CL
- Authors: Priyanshu Mohanty, Sarthakbrata Halder, Srikant Panda, Afshin Oroojlooy, Murari Mandal
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Demonstrates a systematic, reproducible attack surface (linguistic style rewrites) achieving up to 80% ASR across multiple frontier models. Includes mechanistic analysis of why style disrupts refusal mechanisms and a practical defense method.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet