Simulated control evals move agent safeguards closer to deployment
Redwood Research introduced simulated control evaluations, which replay real task context while an LLM simulates tool responses instead of executing attacks. On LinuxArena, the authors report 3.6 percentage points of safety-prediction RMSE at a 5% audit budget versus 10.1 for BashArena, with simulation explaining 78–85% of between-environment variation across key control metrics.
Why it matters: Control evaluations run in generic sandboxes can miss deployment-specific code, services and attack opportunities. A simulator can cheaply test monitors against more realistic context without giving an attack policy live access, but it shifts trust toward…