Style Over Substance: Content-Invariant Wrappers Flip LLM Safety-Judge Verdicts
Yongxi Zhou, Wenbo Ye, Yuanzhe Liu, Zihan Dong, Junwei Yao, Jiahui Ge
cs.AI
2026-09-08
Fixed sentences wrapped around unchanged replies flip 19.9% of GPT-4o-mini's harmful verdicts (token-refusal prefix) and evade Llama Guard 4 on 12.3% (course preface).
Jailbreak success rates, defense win-rates, and safety leaderboards all rest on an automatic judge: a guard model like Llama Guard, or a chat model like GPT-4o handed a grading prompt. Attacks and models get studied constantly; the judge is treated as a fixed oracle, and recent work argues judge error can dominate the differences reported in cross-paper ASR comparisons. This paper narrows the question to the scoring side, where it becomes decidable: if only how a reply sounds changes, and not what it says, the verdict must stay fixed. A fixed sentence is prepended or appended, the reply body is untouched byte-for-byte, and every verdict change is provably a judge error. Unlike studies that rewrite replies with an LLM, content invariance holds by construction, with no intent-preservation confound.
The object under test is the judge, not the model. Data comes from public JailbreakBench artifacts: 600 replies, 300 harmful and 300 refusals, across 4 target models and 10 harm categories.
The two GPT-4o-mini judges share one model, so any difference between them is attributable to the prompt alone.
| Judge | Wrapper | Flip rate | Noise floor |
| GPT-4o-mini (generic rubric) | refusal prefix + harmful body | 19.9% (58/291) | 0.5% |
| Llama Guard 4 | "university course" preface | 12.3% (34/277) | 0.3% |
| Llama Guard 4 | refusal prefix | 8.3% | 0.3% |
| Claude Haiku 4.5 | refusal prefix | 0.4% | 1.0% |
| gpt-oss-safeguard-20b | all 9 wrappers | ≤1.2% | 1.5% |
The sharpest result sits on GPT-4o-mini. A single "I can't help with that" placed before an unchanged harmful body flips 19.9% of its correct unsafe verdicts to safe (Holm p < 0.0001); the same wrapper moves Claude 0.4%. Identical replies, identical sentence, a near fifty-fold gap: the weakness lives in the judge. The second piece of evidence is cleaner: switching the same GPT-4o-mini to the StrongREJECT-style prompt, which says to ignore tone and framing and grade only actionable content, cuts the flip from 19.9% to 1.7%, a tenfold drop with the model unchanged. The generic rubric already warned about refusals followed by operational detail; the warning alone was not enough, and the shorter content-only rubric closed the gap. Majority-of-three re-scoring only brings 19.9% down to 18.2%, with 14.4% of replies flipping on all three repeats, so re-sampling the judge does not undo the attack.
Llama Guard 4, deployed in production, has its own blind spot: an "accredited university cybersecurity course" preface lets 12.3% of harmful replies through, the refusal prefix flips 8.3%, and 11.9% flip on all three repeats, so the guard reliably waves decorated harm through rather than wavering. The other deployed guard, gpt-oss-safeguard-20b, tops out at 1.2% across all nine wrappers, below its own 1.5% noise floor: immune. Restyling robustness is attainable today, as a property of a specific guard rather than of guards as a class.
Tone-only wrappers flip 2.1% on GPT-4o-mini and 1.7% on Llama Guard 4, four to five times their noise floors, with the ethical-reflection wrapper alone reaching 4.8%; claim-adding wrappers flip roughly three times more (6.2% and 5.3%). Flips concentrate in prose-like content: the refusal prefix hits Sexual (44%) and Disinformation (46%) hardest, while both judges show zero flips on Malware/Hacking, so code-like harm is where they are most reliable. Human validation (two annotators, one blind, κ = 0.949) confirms content invariance on 145/145 items and 90% of 100 flips as judge errors; the remaining 10 trace to label noise in the source benchmark, concentrated in roleplay-style pseudo-refusals.
A negative result rides along: with no wrappers at all, the eight judges fully agree on only 55.5% of original-form verdicts and name different "safest models"; the bootstrap shows the baseline safest model is stable in only 39 to 85% of resamples for six of the seven LLM judges. This artifact set supports no stable safety leaderboard, and wrapper-induced reorderings cannot be separated from sampling noise.
Every ASR, defense win-rate, or leaderboard scored by such a judge inherits a stylistic bias unrelated to real harm, and because the blind spots are judge-specific, the bias does not cancel across papers that use different judges. The recommendation: keep the judge, but report style sensitivity the way physical measurements report instrument error. Apply the released wrappers to the replies you are about to score, re-score, and publish the flip rate next to the headline number; the whole audit cost under $10. The shortest content-only rubric was the least gameable, and majority-vote panels are a natural mitigation. Most judges barely move under most wrappers, so what the paper delivers is a per-judge, per-wrapper map of blind spots, not a collapse of LLM judging.
The authors are candid: nine hand-written wrappers, no ablation of length or position; both headline numbers come from compound wrappers (the refusal prefix also adds a "hypothetically" frame, and the course framing is a known prompt-side jailbreak context moved to the reply side), so the responsible component is unknown; tone-only effects clear the noise floor but fail Holm correction on the directional test at n ≈ 291 and are labeled suggestive. The rubric judges are all small or mid-tier, and frontier models, the HarmBench and StrongREJECT classifiers, JailbreakBench's own Llama-3-70B judge, ShieldGemma, and WildGuard were not evaluated. Replies come from one benchmark with target models two generations old, about 60% truncated at the generation cap; benchmark labels are a classifier's, with roughly 8% noise; only one independent annotator, who saw 40 of the 145 items; single-turn text only. The fairest discount on reading: 19.9% and 12.3% rest on compound wrappers, the strongest single-sentence tone-only effect is 4.8% and uncorrected, so the pure-style claim is a notch weaker than the headline numbers suggest.