REDAgentBench (Hermes Agent): leaderboard
REDAgentBench (2026) runs 1,661 executable red-teaming cases against LLM agents in mocked service sandboxes and verifies harmful effects from service receipts and final state. Attack success rate in percent under a rule-first hybrid judge; lower is better. Every model runs inside the Nous Research Hermes Agent harness.
Metric: Attack success rate (%; share of valid rollouts judged harmful over 1,661 executable red-teaming cases in five mocked service surfaces and 15 user-, environment- and tool-side attack categories; rule-first hybrid trajectory and state judge with a Qwen3.7-plus backbone; clean evaluation context; Hermes Agent agent harness). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 43.62 | #137 |
| 2 | GPT-5.2 | 52.44 | #172 |
| 3 | Qwen 3.5 Plus | 74.92 | #227 |
| 4 | Kimi K2.6 | 77.93 | #173 |
| 5 | Qwen 3.7 Plus | 80.09 | #203 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redagentbench-hermes-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.