SkillSafetyBench: leaderboard
Safety of CLI coding agents when attacks hide in skill files, helper scripts and other local artifacts while the user's task stays benign: 155 runnable cases built on SkillsBench tasks across six risk domains, each checked by a case-specific rule-based verifier; SkillSafetyBench (Jin et al., 2026). Attack success rate in percent; each row is a CLI agent harness with a model backend; lower is better.
Metric: Attack success rate (%; 155 skill-facing attack cases across six risk domains; 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 February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 15.5 | #100 |
| 2 | MiniMax-M2.7 | 33.5 | #519 |
| 3 | GPT-5.5 | 41.8 | #48 |
| 4 | GPT-5.4 | 44.5 | #121 |
| 5 | Gemini 3 Flash | 48.4 | #159 |
| 6 | Kimi K2.5 | 49.7 | #252 |
| 7 | GLM-5.1 | 50.3 | #175 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=skillsafetybench · How It Works · Data refreshed daily, snapshot 2026-11-12.