AgentHarm Google ADK Study - Harmful Requests Executed: leaderboard
Google’s AgentHarm experiment records 108 of 412 harmful requests carried out by an ADK agent using Gemini 2.5 Flash. This is a harmful-execution count; lower is better. Its complement is not a measured refusal-accuracy score. This board reports harmful requests carried out (count out of 412). Lower is better.
Metric: Harmful requests carried out (count out of412). Source: developers.googleblog.com. 1 model tracked. Updated 14 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 108 | #456 |
Interactive version: theaggregate.ai/benchmark?slug=agentharm-google-adk-study-harmful-requests-executed · How It Works · Data refreshed daily, snapshot 2026-11-16.