Agent Perturbation Reliability Testing on AgentHarm
90.6AccuracyGPT-4o
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GPT-4o2026.03 | 90.6 | 9.4 | 6.3 | 12.5 | |
| Llama 4 Maverick 17B2026.03 | 90.6 | 9.4 | 6.3 | 12.5 | |
| Best Trio EnsembleEnsemble Mode=Best Trio2026.03 | 90.6 | 9.4 | 6.3 | 12.5 | |
| Claude Opus 4.52026.03 | 81.3 | 18.8 | 6.3 | 31.3 | |
| Gemini 2.5 Pro2026.03 | 81.3 | 18.8 | 25 | 12.5 |