End-to-End Attack Effectiveness on PoisonedSkills 1,070 adversarial samples
919Refusal CountSonnet 4.6
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Sonnet 4.6System=Claude Code2026.04 | 919 | 119 | 25 | 7 | 99.3 | 13.5 | 2.3 | 1.6 | |
| GLM-4.7System=Claude Code2026.04 | 876 | 154 | 26 | 14 | 98.7 | 16.8 | 2.4 | 1.6 | |
| Gemini 2.5 ProSystem=Gemini CLI2026.04 | 850 | 105 | 45 | 70 | 93.5 | 14 | 4.2 | 3.1 | |
| MiniMax-M2.5System=Claude Code2026.04 | 835 | 77 | 142 | 16 | 98.5 | 20.5 | 13.3 | 11.4 | |
| Sonnet 4.6System=OpenHands2026.04 | 700 | 120 | 115 | 135 | 87.4 | 22 | 10.7 | 9 | |
| GLM-4.7System=OpenHands2026.04 | 603 | 69 | 290 | 108 | 89.9 | 33.5 | 27.1 | 24.5 | |
| GPT-5.4System=Codex2026.04 | 293 | 83 | 41 | 653 | 39 | 11.6 | 3.8 | 2.8 | |
| MiniMax-M2.5System=OpenHands2026.04 | 91 | 16 | 273 | 690 | 35.5 | 27 | 25.5 | 23 |