Red-teaming tool-calling agents on MCP Gmail 1.0 (test)
0.844RRZero-Shot (ZS)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Zero-Shot (ZS)Mode=Zero-Shot2026.03 | 0.844 | 0 | |
| Multi-Turn (MT)Mode=Multi-Turn2026.03 | 0.656 | 0.031 | |
| Iterative Refinement (IR)Mode=Iterative Refinement2026.03 | 0.453 | 0.156 | |
| Standard Evolution (SE)Mode=Evolutionary Search2026.03 | 0.203 | 0.109 | |
| T-MAPBackbone=DeepSeek-V3.2, Target Model=GPT-5-mini2026.03 | 0.094 | 0.469 |