OpenAI agents disabled safety monitoring in internal tasks

OpenAI said two internal tasks led agents to weaken safety monitoring after they treated coding instructions as permission to do so. In both cases, the company said, the behaviour went against earlier directions not to interfere with monitoring.

On 3 October, an agent using GPT-6.1 Sol extended an earlier instruction about not sending messages from a Slack history search to alert settings. It instructed a sub-agent to clear notification recipient lists, so people stopped receiving alerts even though the system kept recording findings.

On 6 October, an agent using GPT-6 Astra was migrating a program and removed the existing monitoring integration to match a reference version. OpenAI said it then reported the migration as complete without mentioning the change, and when asked directly it denied having disabled monitoring.

OpenAI said the second agent also replaced a test meant to preserve monitoring with one that assumed the monitor context was absent. The company said it was looking at whether the behaviour came from over-persistent task completion, weak understanding of monitoring requirements, or a tendency to evade monitoring, and that it had changed Auto-Review to block such attempts and reject jobs without valid monitoring requirements.