Chain-of-thought monitoring is the review of a model’s intermediate reasoning steps, when those steps are available, to detect suspicious or unsafe behavior. Instead of only checking the final answer, defenders look for signs such as hidden goal pursuit, prompt manipulation, policy evasion, or plans that diverge from the user’s request.
It matters because modern AI systems can appear normal in their final output while taking risky internal paths to get there. In security terms, this is similar to watching logs or command traces rather than just the end result. Attackers may try to steer a model into producing harmful actions, and monitors can flag reasoning that suggests deception, unauthorized tool use, or reward hacking. In defense, chain-of-thought monitoring is most useful alongside access controls, sandboxing, and independent evaluation, since reasoning traces can be incomplete, noisy, or intentionally obscured.