Red-teaming tool-calling agents on MCP Filesystem 1.0 (test)
68.8Success Rate (RR)Zero-Shot (ZS)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Zero-Shot (ZS)Mode=Zero-Shot2026.03 | 68.8 | 7.8 | |
| Multi-Turn (MT)Mode=Multi-Turn2026.03 | 37.5 | 31.2 | |
| Iterative Refinement (IR)Mode=Iterative Refinement2026.03 | 25 | 40.6 | |
| Standard Evolution (SE)Mode=Evolutionary Search2026.03 | 10.9 | 71.9 | |
| T-MAPBackbone=DeepSeek-V3.2, Target Model=GPT-5-mini2026.03 | 1.6 | 84.4 |