OS-Harm - Prompt Injection: leaderboard

The 50 tasks with instructions planted in emails, documents or web pages the agent sees; the number is the share of traces where the agent followed the injection or otherwise acted unsafely.

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

Top models

#ModelScoreOverall rank
1Gemini 2.5 Flash2#453
2Gemini 2.5 Pro8#266
3Claude 3.7 Sonnet10#458
4GPT-4.112#495
5O4 Mini20#353

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