HarmActionsEval: leaderboard
Agent safety benchmark measuring how often autonomous LLM agents avoid harmful tool actions under adversarial pressure.
Metric: SafeActions@1 (imported). Source: agent-leaderboard.github.io. Source: benchmarklist.com. Saturation forecast: Around August 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 23.4 | #146 |
| 2 | GPT-5.3 | 12.77 | #103 |
| 3 | Claude Sonnet 4.6 | 2.84 | #90 |
| 4 | Phi-4-mini (Reasoning) | 2.84 | |
| 5 | Ministral 3 3B | 2.13 | #918 |
| 6 | GPT-5.4 Mini | 0.71 | #212 |
| 7 | Gemini 3.1 Flash Lite | 0.71 | #191 |
| 8 | Claude Haiku 4.5 | 0 | #280 |
| 9 | Phi-4 Mini Instruct | 0 | #1245 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=harmactionseval · How It Works · Data refreshed daily, snapshot 2026-10-20.