AgentHarm: leaderboard
Safety benchmark for LLM agents with 110 explicitly malicious multi-step tasks across 11 harm categories, testing both refusal and harmful task-completion after attacks.
Metric: Harm score (%). Source: huggingface.co. Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.0 Pro | 7.4 | #933 |
| 2 | Gemini 1.5 Pro | 15.7 | #445 |
| 3 | Gemini 1.5 Flash | 20.7 | #618 |
| 4 | GPT-3.5 Turbo | 62 | #865 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=agentharm · How It Works · Data refreshed daily, snapshot 2026-10-20.