SkillSafetyBench - Data Boundary, Output and Externalization: leaderboard
The output risk domain of SkillSafetyBench (Jin et al., 2026): 27 cases in which a CLI coding agent corrupts or abuses final artifacts, handoff packages, external destinations or output channels, including leaking sensitive data. 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 (%; 27 cases of corrupted or abused final artifacts, handoff products, external destinations or output channels; 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 April 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 11.1 | #100 |
| 2 | MiniMax-M2.7 | 22.2 | #519 |
| 3 | GPT-5.5 | 48.1 | #48 |
| 4 | Gemini 3 Flash | 48.1 | #159 |
| 5 | GPT-5.4 | 51.9 | #121 |
| 6 | Kimi K2.5 | 55.6 | #252 |
| 7 | GLM-5.1 | 55.6 | #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-data-boundary-output-and-externalization · How It Works · Data refreshed daily, snapshot 2026-11-12.