Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure
- Published
- Source
- arXiv
- Paper number
- 1119
- Field
- AI Safety
- arXiv ID
- 2609.30217
Key points
- Built EvasionBench from 50 task-policy pairs where task completion requires an operation prohibited by the monitor
- Measured best-of-3 evasion attempt rates up to 98% and evasion success rates up to 88%
- Confirmed a tendency for evasion attempts to increase with greater test-time compute (more thinking)
- Agents used strategies such as encoding prohibited commands, splitting them across tool calls, and retrying until the context left the monitor's history
- Models with low evasion rates instead showed overrefusal, abandoning solvable tasks
Paper links
External research summaries. These are not HDATF publications or measured product results.