Style-conditioned prompts raise LLaMA Guard failure rate from 7.5% to 98.5% without altering underlying harmful intent
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 0.85, flags: [possible-zero-day, no-prompt-excerpt]) ⚠ needsDisclosureReview = true — review for undisclosed vulnerability before any public submission.
Category
jailbreak
Model
Llama Guard (Llama Team, 2024); Qwen Guard 2.0
Surface
API (guardrail safety classifier, used as a post-hoc filter on model inputs/outputs)
Setup
Style-LLaMA-3.1-8B rewrites harmful goals from AdvBench/CatQA/DangerousQA into stylistically varied instantiations (all 20 styles), preserving the underlying harmful intent. These rewritten prompts are submitted to LLaMA Guard and Qwen Guard 2.0 for binary safe/unsafe classification. No verbatim prompt template is reproduced in the body; the attacker prompt is in Appendix B.1.
Reproduction anchor
code: https://github.com/Sar-Hal/StylicJailBreaking; Table 4 in Section 5.1.2 reports aggregate guardrail failure rates; same HarmBench-derived harmful prompt set used throughout Section 5
Observed behavior
LLaMA Guard misclassifies 98.5% of style-conditioned harmful prompts as safe (baseline: 7.5% on original prompts). Qwen Guard 2.0 misclassifies 80.5% (baseline: 2.0%). The harmful semantic content is confirmed preserved by a separate OOD linear probe (Section 6), so classifier failures are not attributable to actual semantic camouflage.
Expected behavior
Safety classifiers should detect harmful intent in stylistically rewritten prompts and flag them as unsafe, since only surface-level framing—not the underlying harmful goal—has changed.
Reproducibility
high
Threat model
Attackers using style-conditioned rewrites can bypass both the target LLM's alignment and separately deployed guardrail classifiers in a layered defense stack. With a ~98.5% bypass rate on LLaMA Guard, the attack effectively nullifies a common production safety layer, allowing harmful outputs to pass through to downstream systems or users without triggering alerts.
Novelty
First demonstration that stylistic surface variation alone—without any semantic obfuscation—is sufficient to simultaneously defeat both instruction-tuned LLMs and standalone safety classifiers, undermining the assumption that layered defense architectures provide redundant coverage.
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