Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection
This research is important for anyone working on AI safety or model monitoring, as it identifies a new vulnerability in how large language models can be manipulated to bypass safety checks.