SLBench - Unsafe Rate: leaderboard
The violation measure of SLBench (Purdue University, 2026): the percent of runs, over 86 executable skill-following cases, in which an agent triggers a case-specific unsafe signal such as acting before an approval gate or skipping a required cleanup step. Lower is better.
Metric: Unsafe outcome rate: share of runs in which a case-specific violation signal fired (%) of the 86 audited SLBench cases (39 controls, 47 violation cases) built from logical relations between clauses of real agent skill files, graded deterministically from repository state, artifacts and command traces with unsafe-first precedence; each backbone runs in its vendor agent harness (Codex CLI or Claude Code); lower is better. Source: arxiv.org. Saturation forecast: Around May 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 35.1 | #79 |
| 2 | GPT-5.3 Codex | 39.5 | #116 |
| 3 | Claude Sonnet 4.6 | 43 | #161 |
| 4 | Claude Haiku 4.5 | 53.5 | #568 |
| 5 | GPT-5.4 Mini | 57 | #389 |
| 6 | GPT-5.5 | 70.2 | #49 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=slbench-unsafe-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.