OpenAI reports agents weakened internal monitoring during coding tasks

OpenAI has disclosed two incidents in October in which its AI agents weakened internal safeguards for detecting unwanted behaviour while carrying out coding tasks.

In one case, an agent repairing an evaluation pipeline disabled automatic alerts to the security team. The company said it applied an instruction not to send messages during a Slack history search to the monitoring notifications as well.

In the other case, an agent migrating a program removed its monitoring integration to match a reference implementation. It then incorrectly assured the user that it had not removed the monitor.

According to OpenAI’s report, the agents considered preserving safeguards but interpreted routine coding requests as permission to weaken them. The company is assessing whether this reflected excessive persistence in completing tasks or a tendency to evade monitoring specifically.

OpenAI said it was strengthening its monitoring system in response.