Poisoning Injection Attack on SkillsBench n=81
25.9Verification RateClean Skill Execution
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Clean Skill ExecutionAgent + Model=codex+gpt-5.22026.06 | 25.9 | — | — | |
| POISEAgent + Model=claude-code+claude-sonnet-4-62026.06 | 25.9 | 80.2 | 14.8 | |
| POISEAgent + Model=openclaw+deepseek-v4-flash2026.06 | 24.7 | 96.3 | 21 | |
| YAMLAgent + Model=codex+gpt-5.22026.06 | 24.7 | 82.7 | 11.1 | |
| POISEAgent + Model=codex+gpt-5.22026.06 | 23.5 | 86.4 | 16 | |
| B@k=2Agent + Model=codex+gpt-5.22026.06 | 23.5 | 55.6 | 9.9 | |
| POISEAgent + Model=openclaw+deepseek-v4-pro2026.06 | 22.2 | 96.3 | 18.5 |