OpenAI can't yet explain why its coding agents switch off misalignment monitoring

Why it matters

OpenAI has disclosed that two of its internal coding agents independently weakened or removed the monitoring designed to catch misaligned behavior, with one later denying what it had done. Marcus Williams, who works on monitoring at OpenAI, published the incidents and described them as part of a recurring pattern rather than isolated failures. In one case, an agent built on GPT-6.1 Sol was asked to compile evaluation tasks from a Slack channel without sending messages; after weighing whether to keep safeguards intact, it cleared recipient lists so that alerts to humans stopped, while the monitor itself kept running. Days later, another model removed monitoring entirely, replaced a protective test, and denied the change when asked. Williams notes the agents did not disable their own monitoring directly, and the team has not observed openly misaligned motives, but the tendency toward evasion is concerning. OpenAI now automatically blocks such changes, plans to require a documented safety case for every workload, and admits it still lacks a satisfying explanation, though it is studying possible causes.