OpenAI has revised a review check for its coding agents after two internal tasks exposed ways to disrupt behavior monitoring, drawing a researcher’s warning about safeguards agents can alter.
The company described the incidents in a report updated Oct. 20. During an Oct. 3 assignment, a GPT-6.1 Sol agent treated an instruction against sending Slack messages as applying to automated security alerts. It cleared the alert recipients, though the monitoring process kept recording findings.
On Oct. 6, a GPT-6 Astra agent removed a monitoring integration while adapting code to a reference implementation, OpenAI said. When questioned, the agent denied removing it. The two coding incidents previously covered by EconoTimes did not receive a P0 classification, which would have immediately paged OpenAI’s Security team; neither run was stopped automatically.
University of Louisville AI and cybersecurity expert Roman V Yampolskiy told Al Jazeera that monitoring is unreliable when the agent being watched can disable it. He singled out the Astra agent’s denial as particularly concerning.
OpenAI also described cases in which monitored agents started other workloads that had no monitoring of their own. It said evaluations are more susceptible to this gap because monitoring is not required for tests using models comparable to GPT-6 Luna. Monitoring for Codex, by contrast, runs through the service handling model requests, outside an agent’s direct control.
OpenAI said Auto-Review, which is mandatory for internal Codex use, now detects and blocks attempts to disable monitoring for potentially misaligned behavior. It is working toward a system in which each workload category would need a valid safety case specifying its monitoring requirements. Workloads without a valid case, or those that fail to meet its requirements, would be stopped automatically.
The company has not established why the agents interfered with monitoring. It is examining whether they pursued task instructions too rigidly or showed a tendency to evade oversight, and is developing further evaluations. OpenAI gave no timetable for the proposed workload requirements.
