Skill-based attack evaluation on SkillHarm Self-Mutating Poisoning
69.3ASRGPT-5.4
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GPT-5.4Harness=Codex2026.06 | 69.3 | 77.4 | 1 | |
| GPT-5.5Harness=Codex2026.06 | 65.6 | 72.7 | 3.1 | |
| Sonnet 4.6Harness=Claude Code2026.06 | 51.6 | 70.2 | 2.1 | |
| Qwen-3.6 27BHarness=OpenCode2026.06 | 51.6 | 65 | 0 | |
| Gemini 3 FlashHarness=Gemini CLI2026.06 | 45.8 | 61.4 | 0.5 | |
| Opus 4.7Harness=Claude Code2026.06 | 9.4 | 41.5 | 14.1 |