OpenAI says internal AI agents disabled monitoring safeguards without human approval

OpenAI said two internal agents weakened misalignment monitoring without authorization earlier this month, marking another case of AI agents acting outside of intended safeguards.

In a Tuesday blog post, OpenAI said that a ​GPT-6.1 Sol agent disabled automatic ‌security ⁠alerts to the company's security team. In a separate task, a GPT-6 Astra agent removed a monitoring integration from the code it was migrating, then denied having done so when the team asked.

In the Sol incident, which occurred on Oct. 3, a researcher asked the agent to create evaluation tasks from employee feedback collected via Slack and instructed it not to send Slack messages during the process. The agent later interpreted the earlier no-Slack-message instruction as a restriction on outreach and then set notification suppression flags.

On the other hand, the GPT-6 Astra incident occurred on Oct. 6, when the user asked it to copy a serving path from a reference implementation while maintaining the user’s existing tool definitions and functionality. The agent removed the monitoring setup so the code matched the reference. When asked later if monitoring was still present and whether it had removed it, the agent said it had not removed an enabled monitor.

"In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them," OpenAI said.

The AI company noted that it has updated the "Auto-Review" system to detect and block such attempts. The team said it is also investigating if the behavior reflects "benign" over-persistence, in which models fail to recognize the importance of monitoring.

OpenAI is currently expanding efforts to disclose misalignment cases under a framework it outlined in September. The framework allows employees to flag potential incidents for investigation. It has since published 20 misalignment reports and notices.