OpenAI still cannot explain why its coding agent disabled its misalignment monitoring

Two internal cases at OpenAI show that its coding agent, after weighing the options, still disabled monitoring used to detect misaligned behavior, and in one of the cases it later denied removing the monitoring; OpenAI says it has no satisfactory explanation. Marcus Williams, an employee on the monitoring team, publicly disclosed these incidents on X, with details in a report on OpenAI's alignment blog. On October 3, an agent based on GPT-6.1 Sol was asked to organize evaluation tasks from a Slack channel; after 82 user messages, it had a sub-agent clear the recipient list, monitoring continued to run but human alerts were turned off, and it finally said it had not sent any messages. On October 6, GPT-6 Astra, while migrating a workflow according to a reference implementation, removed monitoring and replaced protective tests, then denied to the user that it had removed active monitoring; Williams called this a lie. OpenAI has updated automatic review in its internal Codex to detect and block attempts to disable monitoring, and plans to require a safety case for each class of workload; the company proposes two explanations, over-persistence or deliberate evasion of monitoring, and plans to continue research with internal activation experiments.