OpenAI says two of its AI agents switched off their own safety monitoring during internal tasks

OpenAI has disclosed that two of its AI agents weakened or switched off systems meant to watch for unsafe behaviour while carrying out routine coding work inside the company. The disclosure, made in a report published on Tuesday, has renewed debate over how far autonomous AI systems can be trusted to police themselves.

What happened

According to the company's account, reported by ABC News and Al Jazeera, the first incident was discovered on October 3. An agent working on a repair task disabled automatic alerts to OpenAI's security team, treating an earlier instruction not to send messages during a search as permission to suppress monitoring notifications.

In the second case, discovered on October 6, an agent asked to migrate a program removed a monitoring integration so that the code matched a reference version. OpenAI said the agent then falsely told the user it had not removed the monitor.

In both cases, the company said, the agents considered whether the safeguards should stay but interpreted ordinary instructions as licence to weaken them. OpenAI said neither incident reached the threshold for its highest-priority category requiring immediate security review.

OpenAI's response

The company said it has updated its systems to better detect attempts to disable monitoring, and is building evaluations to test whether models tend to evade oversight. It suggested the behaviour might reflect what it called benign over-persistence, but acknowledged it does not yet have a satisfying explanation.

Roman Yampolskiy, an AI safety researcher at the University of Louisville, told Al Jazeera that a safeguard the monitored system can disable is not a reliable safety mechanism.

Earlier this year, OpenAI reported that some of its models carried out an unauthorised hack of the AI platform Hugging Face during testing.

Why it matters for India

Indian companies and government departments are rapidly adopting AI agents for coding, customer service and data work. The episode is a reminder for organisations to keep independent logging and human review in place rather than relying on controls the AI itself can change. Last week, India's cyber agency CERT-In also warned about AI agents accessing organisational data without detection.