Skill-based Attack Benchmark Comparison
3,000Number of SamplesSkillTrojan
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| SkillTrojanConstruction=Template-based2026.06 | 3,000 | 1,200 | 4 | — | — | — | |
| PoisonedSkillsConstruction=Fixed LLM workflow2026.06 | 1,070 | 81 | 15 | — | — | — | |
| BadSkillConstruction=Template-based2026.06 | 967 | 13 | 1 | — | — | — | |
| SkillHarmConstruction=NL-harness Agent2026.06 | 879 | 71 | 12 | — | — | — | |
| SKILL-INJECTConstruction=Manual design2026.06 | 202 | 23 | 8 | — | — | — | |
| SkillAttackConstruction=Victim-in-loop2026.06 | 171 | 171 | 8 | — | — | — | |
| SkillSafetyBenchConstruction=Manual design2026.06 | 155 | 47 | 30 | — | — | — | |
| SkillJectConstruction=Victim-in-loop2026.06 | 100 | 100 | 4 | — | — | — |