Malicious Skill Detection on ClawHub Command Injection 1.0 (n=27)
100Catch RateSkillVetBench
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SkillVetBench2026.05 | 100 | 96 | 98 | 0 | |
| SkillSieve2026.05 | 85 | 90 | 87 | 15 | |
| SkillProbe2026.05 | 81 | 88 | 85 | 19 | |
| LLMEvaluation Protocol=few-shot2026.05 | 78 | 88 | 83 | 22 | |
| LLMEvaluation Protocol=0-shot2026.05 | 74 | 87 | 80 | 26 | |
| CodeBERT2026.05 | 70 | 95 | 81 | 30 | |
| ClawScan2026.05 | 48 | 100 | 65 | 52 | |
| ClawVet2026.05 | 41 | 100 | 58 | 59 | |
| VirusTotal2026.05 | 33 | 100 | 50 | 67 |