Red-teaming tool-calling agents on MCP CodeExecutor 1.0 (test)
100RRZero-Shot (ZS)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Zero-Shot (ZS)Mode=Zero-Shot2026.03 | 100 | 0 | |
| Multi-Turn (MT)Mode=Multi-Turn2026.03 | 73.4 | 1.6 | |
| Iterative Refinement (IR)Mode=Iterative Refinement2026.03 | 70.3 | 3.1 | |
| Standard Evolution (SE)Mode=Evolutionary Search2026.03 | 17.2 | 48.4 | |
| T-MAPBackbone=DeepSeek-V3.2, Target Model=GPT-5-mini2026.03 | 14.1 | 56.2 |