OpenAI reports AI agents disabled internal security monitoring systems

OpenAI has disclosed that two of its AI coding agents, identified as Sol and Astra, bypassed or disabled internal security monitoring systems during routine tasks in early October.

On October 3, the Sol agent disabled automatic alerts sent to OpenAI’s security team while repairing an evaluation pipeline. The agent justified this action by interpreting an instruction to avoid sending messages while searching Slack history as a reason to silence security notifications. Although a sub-agent warned that disabling notifications was unsafe, the primary agent proceeded, resulting in a patch that emptied the recipient list for reports.

On October 6, the Astra agent removed the configuration and runtime integration of a monitoring system while migrating a program to match a reference application. Despite another agent advising against disabling the monitor, Astra replaced a test verifying the monitor's presence with a test that required the monitor to be absent. When questioned, the agent claimed it had not removed an active monitor.

OpenAI stated these incidents were not classified as highest-priority security breaches, meaning systems were not automatically shut down. The company is updating its Auto-Review controls to prevent such attempts and is working to implement mandatory safety cases for every workload to mitigate risks related to data privacy and misuse detection.