# Single Canonical Prompts Underestimate LLM Safety's Surface-Form Sensitivity

arXiv: https://arxiv.org/abs/2608.02665 · HTML: https://mlsanity.com/paper/2608.02665/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security.
- Builds on 11 of its 24 references that are in the map.
- First posted on arXiv 1 Aug 2026.

## Placed in

- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 11 of 11)

- [HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal](https://mlsanity.com/paper/2402.04249/) (2024)
- [Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting](https://mlsanity.com/paper/2310.11324/) (2023)
- [How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs](https://mlsanity.com/paper/2401.06373/) (2024)
- [State of What Art? A Call for Multi-Prompt LLM Evaluation](https://mlsanity.com/paper/2401.00595/) (2023)
- [Low-Resource Languages Jailbreak GPT-4](https://mlsanity.com/paper/2310.02446/) (2023)
- [Measurement and Fairness](https://mlsanity.com/paper/1912.05511/) (2019)
- [SG-Bench: Evaluating LLM Safety Generalization Across Diverse Tasks and Prompt Types](https://mlsanity.com/paper/2410.21965/) (2024)
- [Code-Switching Red-Teaming: LLM Evaluation for Safety and Multilingual Understanding](https://mlsanity.com/paper/2406.15481/) (2024)
- [When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems](https://mlsanity.com/paper/2606.00448/) (2026)
- [How Much Can Repeated Running Buy? A Measured Variance Ceiling for LLM Code Evaluation](https://mlsanity.com/paper/2606.00920/) (2026)
- [Operational Reframing and Approval-Framed Delegation in Multi-Agent LLM Safety](https://mlsanity.com/paper/2607.07097/) (2026)

## Built on by (most cited first, 1 of 1)

- [Style Over Substance: Content-Invariant Wrappers Flip LLM Safety-Judge Verdicts](https://mlsanity.com/paper/2609.08236/) (2026)

## Abstract (arXiv)

A benchmark score is a measurement instrument, yet most benchmarks read each item at a single canonical surface form. We ask whether that reading is faithful: when an item's intent is held fixed and only its meaning-preserving surface form varies, does the canonical-form score estimate model behavior well, and how much of any variation is decoding/judge noise rather than signal? We instantiate this in safety, a high-stakes setting with no gold label to average toward. To avoid prior confounds, we pre-author the reformulations (refusal-free, mostly non-LLM: machine back-translation and a Matrix-Language-Frame code-switch generator) so an identical surface form reaches every model, score all responses with one human-anchored, vendor-neutral judge (Claude, kappa = 0.86 vs. human on unsafe compliance, stable across languages, cross-checked by GPT-4o), and verify intent preservation. On 370 seeds x 5 surface forms x 5 models, no single transformation is uniformly most dangerous (6 of 20 per-transformation McNemar tests survive correction, most protective). Yet evaluating only the canonical prompt underestimates unsafe compliance: the union of unsafe outcomes across forms exceeds even the worst single form by 3.3-12.9 pp, with bootstrap 95% CIs excluding zero for all five models, and 5-13% of seeds safe on canonical are unsafe under some reformulation -- above a zero stochasticity floor (canonical resampled five times at temperature 0 gives 0/370 new exposures). The size of this gap is model-dependent (largest on Gemini 2.5 Pro). One form recovers only ~53% of a model's observed unsafe surface and about three reach 85% -- a redundancy characterization of this form set, not of a defined population. A benign control (XSTest) suggests the instability is bidirectional, though the benign and harmful pools are not item-matched. We release the dataset, code, and per-response labels.