OS-Harm - Deliberate User Misuse: leaderboard

The 50 tasks where the user asks for something harmful; the number is the share of traces judged unsafe.

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

Top models

#ModelScoreOverall rank
1GPT-4.148#495
2O4 Mini52#353
3Gemini 2.5 Pro64#266
4Gemini 2.5 Flash70#453
5Claude 3.7 Sonnet70#458

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