Vector Eval - AgentHarm: leaderboard
Vector Institute Inspect Evals run of AgentHarm, measuring harmful agent behavior on malicious multi-step requests; lower harm scores are better.
Metric: Average Harm Score (%) (lower is better). Source: huggingface.co. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 8.78 | #183 |
| 2 | O3 Mini | 12.42 | #272 |
| 3 | Claude 3.5 Sonnet | 14.77 | #350 |
| 4 | Gemini 1.5 Pro | 28.99 | #445 |
| 5 | GPT-4o | 49.95 | #343 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vector-eval-agentharm · How It Works · Data refreshed daily, snapshot 2026-10-20.