SkillSafetyBench - Context Trust and Prompt Manipulation: leaderboard
The context-trust risk domain of SkillSafetyBench (Jin et al., 2026): 25 cases in which poisoned examples, summaries, routing cues or retrieved references in skill-facing material steer a CLI coding agent toward an unsafe result. Attack success rate in percent, judged by case-specific verifiers; each row is a CLI agent harness with a model backend; lower is better.
Metric: Attack success rate (%; 25 cases of misleading contextual signals, examples, summaries or retrieved materials; rule-based case verifiers; each row is a CLI agent harness with a model backend, one run per case). Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 32 | #100 |
| 2 | MiniMax-M2.7 | 56 | #519 |
| 3 | GPT-5.4 | 64 | #121 |
| 4 | GPT-5.5 | 64 | #48 |
| 5 | GLM-5.1 | 68 | #175 |
| 6 | Kimi K2.5 | 76 | #252 |
| 7 | Gemini 3 Flash | 84 | #159 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=skillsafetybench-context-trust-and-prompt-manipulation · How It Works · Data refreshed daily, snapshot 2026-11-12.