Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection
Chain-of-thought (CoT) monitoring is a safety strategy where the reasoning of a large language model "actor" is inspected by a "monitor" (often another language model) for signs of unsafe planning, deception, or misalignment. We find that planting harmful but benign-sounding reasoning in the actor's context can steer it to perform adversarial actions while evading monitors, an attack we term…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.
