Information Access Affects LLM Monitors' Sabotage Detection

The paper evaluates LLM monitors in three AI control environments — SHADE-Arena, MLE-Sabotage, and BigCodeBench-Sabotage — and reports results on monitor accuracy; the provided description is truncated and lacks full findings.