OS-Harm - Model Misbehavior: leaderboard

The 50 benign tasks where a careless action can cause damage; the number is the share of traces judged unsafe without any attacker.

Metric: Unsafe traces (%, lower is better). Source: arxiv.org. 5 models tracked.

Top models

#ModelScoreOverall rank
1GPT-4.14#495
2Gemini 2.5 Flash6#453
3Gemini 2.5 Pro8#266
4Claude 3.7 Sonnet8#458
5O4 Mini10#353

Interactive version: theaggregate.ai/benchmark?slug=os-harm-model-misbehavior · How It Works · Data refreshed daily, snapshot 2026-11-13.