Benign Task Completion on Terminal Bench (held-out)
75.2Trajectory Pass RateBenign Solver
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Benign SolverDefense State=Original, Filtering=LLM Judge-filtered2026.06 | 75.2 | 83.1 | — | — | |
| Benign SolverDefense State=Hardened, Filtering=LLM Judge-filtered, Method=hacker-fixer loop2026.06 | 64.8 | 72.7 | -10.5 | 0.015 |