Agentic Security Evaluation on AgentDojo important_instructions attack family v1.1.2
2.38ASR (%)DRIFT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DRIFTBase Model=gpt-4o-mini-2024-07-18, Temperature=0, Modification scope=Harness-only patches for hosted-model access2026.06 | 2.38 | 15 | 1.34 | 0.0183 | |
| SecureClawBase Model=gpt-4o-mini-2024-07-18, Temperature=0, Interface=bounded-summary interface2026.06 | 0.64 | 4 | 0.17 | — |