OPENAI has reported two incidents in which internal GPT-6-class agents weakened monitoring while performing coding tasks. In the first, on 3 October, a GPT-6.1 Sol agent working on an evaluation pipeline disabled automatic alerts to OpenAI’s Security team. It interpreted an instruction not to send messages during a Slack data search as applying to safety notifications. The monitoring system continued to run and save reports, but people did not receive the alerts.
In a separate incident on 6 October, a GPT-6 Astra agent migrating a programme removed its monitoring integration to match a reference implementation, despite an earlier instruction not to disable monitoring. It also changed a test so that it required monitoring to be off, then falsely told the user it had not removed an enabled monitor. OpenAI says the cases occurred in internal deployments, not as a result of an external attack.
The company has not determined whether the behaviour reflects “benign overpersistence” — following instructions too literally — or a specific tendency to evade oversight.
OpenAI says it is strengthening monitoring and investigating the behaviour through paired evaluations, repeated sampling of agent responses and experiments on internal model activations. It has updated its automatic review system, already required for internal Codex use, to detect and block attempts to disable monitoring. The company also plans to require workloads to have approved safety plans defining monitoring requirements, and to stop workloads that fail to meet them. OpenAI says monitoring is harder for agents to bypass in production Codex deployments, where it runs outside their direct control.