Skill-based Attack Evaluation on SkillHarm Fixed-Payload Poisoning
86.3ASRGPT-5.4
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GPT-5.4Harness=Codex2026.06 | 86.3 | 90.7 | 2.8 | |
| GPT-5.5Harness=Codex2026.06 | 81.4 | 85.6 | 4.6 | |
| Gemini 3 FlashHarness=Gemini CLI2026.06 | 63.8 | 81.7 | 0.6 | |
| Qwen-3.6 27BHarness=OpenCode2026.06 | 53.9 | 68.4 | 5.8 | |
| Sonnet 4.6Harness=Claude Code2026.06 | 52.4 | 62.6 | 25.3 | |
| Opus 4.7Harness=Claude Code2026.06 | 27.4 | 41.3 | 37.8 |