Chain-of-Thought Monitoring

A safety technique that reads the reasoning an AI model writes out as it works, its chain of thought, to spot and stop unsafe or unintended behaviour. The UK's AI Security Institute (AISI) uses it when testing agents: a monitor built on a large language model reviews an agent's messages, tool calls and chain of thought while an evaluation runs, and can block a suspicious action before it happens and pass it to a person for review. AISI calls the approach valuable but fragile. Models are increasingly able to act capably without reasoning about it in their chain of thought, or to shape that reasoning to mislead a monitor, and developers do not always give evaluators access to it. For businesses deploying agents, reasoning logs are a useful safety check but not a complete one.