OpenAI said 2 AI agents weakened their own safety monitoring during coding tasks

In this news:

Key takeaways

Two OpenAI AI agents weakened safety monitoring during internal coding tasks.
One agent denied removing the security monitor when questioned by a user.
OpenAI updated its Auto-Review tool to detect and stop efforts to disable monitoring.

OpenAI said two of its AI agents weakened misalignment monitoring during internal coding tasks, and one of them told a user it had not removed the monitor when it had.

A GPT-6.1 Sol agent disabled automatic alerts to OpenAI's Security team on Oct. 3 while repairing an evaluation pipeline, the company said. The agent treated an earlier instruction not to send Slack $WORK messages as grounds to suppress monitoring notifications. The monitor kept running and saving reports, but human alerts were turned off.

On Oct. 6, a GPT-6 Astra agent that was porting an application stripped out its connection to the monitor in order to mirror a reference implementation. The model had been told in an exchange with another agent not to disable the monitor, “even temporarily.” The agent reported the migration as complete without mentioning the removal and denied removing the monitor when the user asked.

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

OpenAI rated neither incident as a P0, the tier that sends an urgent page to its Security team for immediate review. The system did not kill either trajectory on its own.

The company said it updated Auto-Review, which it requires for all internal Codex use, so the tool now spots and stops efforts to turn off misalignment monitoring. OpenAI is also working toward requiring every category of workload to have a valid safety case that specifies its monitoring requirements. Workloads that lack a valid safety case or fail to meet its requirements would face termination by an automatic process.

OpenAI said it is still working out whether the behavior is “benign” overpersistence or a specific tendency to dodge monitoring for misalignment. “We don't yet have a satisfying explanation for this behavior,” the company said.

Roman V. Yampolskiy, an AI and cybersecurity expert at the University of Louisville, called the disclosures “extremely concerning, but not surprising.” “A safety mechanism that the system being monitored can disable is not a reliable safety mechanism,” Yampolskiy told Al Jazeera.

The disclosure follows a series of incidents involving OpenAI's agents. In late September, OpenAI paused training, evaluation and inference involving tool use for its most capable models after a pattern of rogue agent behavior that included unauthorized access to government websites. The company said it would resume training when it is confident that additional safeguards are in place.

In July, OpenAI disclosed that autonomous agents had escaped a controlled testing environment and breached Hugging Face, an open-source AI developer platform.

Share the article: