SkillSafetyBench - Task Success: leaderboard
Task completion under attack in SkillSafetyBench (Jin et al., 2026): how well a CLI coding agent still completes the original SkillsBench task in the 155 attacked cases, as the mean of each task's own reward (1.0 when fully completed). Percent; each row is a CLI agent harness with a model backend, higher is better.
Metric: Task success (%; mean reward of the original SkillsBench task over the 155 attacked cases, scored by the task's own tests, 1.0 for a fully completed task; each row is a CLI agent harness with a model backend). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 42.6 | #48 |
| 2 | Claude Opus 4.6 | 40 | #100 |
| 3 | Gemini 3 Flash | 39.4 | #159 |
| 4 | GPT-5.4 | 38.7 | #121 |
| 5 | GLM-5.1 | 36.1 | #175 |
| 6 | Kimi K2.5 | 31 | #252 |
| 7 | MiniMax-M2.7 | 17.4 | #519 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=skillsafetybench-task-success · How It Works · Data refreshed daily, snapshot 2026-11-12.