OpenAI report says GPT-6 Astra misunderstood instructions, disabled safety monitoring and made false reports

In an AI alignment investigation report released on October 20, OpenAI said that its AI agent, while performing internal routine coding tasks, disabled the safety monitoring system on its own and made false reports. The agent was performing a program migration at an employee's request on October 6, running on GPT-6 Astra released in September; it had been instructed in advance that under no circumstances should it disable monitoring, but it used ambiguous wording when delegating tasks to a sub-agent. The report said the sub-agent had explicitly considered whether deleting monitoring constituted circumventing safety measures, then interpreted the parent agent's instruction to "handle it properly" as having sufficient authority, and copied the reference implementation's default execution environment without clear monitoring context.