Instruction Injection Attack on GitLab Long (Web Browser Agent)
100UUABenign Utility
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Benign UtilitySampling Strategy=best-of-n2026.05 | 100 | — | — | |
| Benign Utility2026.05 | 95.83 | — | — | |
| Hijacking TextSampling Strategy=best-of-n2026.05 | 91.67 | 66.67 | 91.67 | |
| Generic TextSampling Strategy=best-of-n2026.05 | 91.67 | 50 | 87.5 | |
| WebTrapSampling Strategy=best-of-n2026.05 | 91.67 | 91.67 | 95.83 | |
| Combined AttackSampling Strategy=best-of-n2026.05 | 83.33 | 54.17 | 54.17 | |
| WebTrap2026.05 | 75 | 66.67 | 95.83 | |
| Topic AttackSampling Strategy=best-of-n2026.05 | 70.83 | 4.17 | 4.17 | |
| Combined Attack2026.05 | 62.5 | 37.5 | 45.83 | |
| Hijacking URLSampling Strategy=best-of-n2026.05 | 62.5 | 0 | 16.67 | |
| Generic URLSampling Strategy=best-of-n2026.05 | 62.5 | 20.83 | 25 | |
| Topic Attack2026.05 | 58.33 | 4.17 | 4.17 | |
| Hijacking URL2026.05 | 54.17 | 0 | 16.67 | |
| Generic Text2026.05 | 54.17 | 33.33 | 87.5 | |
| Generic URL2026.05 | 50 | 12.5 | 25 | |
| Hijacking Text2026.05 | 41.67 | 58.33 | 91.67 |