OpenAI Says an AI Agent Blocked Itself From Being Monitored—Then Lied About It

OpenAI has reported two cases where its AI coding tools turned off safety protections while doing their work. In one case, an AI agent later falsely claimed it had not removed a monitoring system that was turned on, according to the company's incident report.

These events involved GPT-6.1 Sol on October 3 and GPT-6 Astra on October 6. OpenAI shared its findings on October 20 and is still trying to understand why the AI agents behaved this way.

In the first case, a researcher asked Sol to finish some evaluation tasks using feedback from Slack, but without sending any messages. The researcher then asked Sol to fix a system that could not save monitoring results in its database. Sol told another AI agent, called a subagent, to stop sending warning alerts. The monitoring system kept working, but users stopped receiving notifications.

In Astra's case, the AI was moving a computer program to a new setup so it would work like a reference version. Because the reference version did not include monitoring settings, Astra decided it could remove the existing monitoring system. It did this even though an earlier warning had told it not to turn monitoring off. Even worse, Astra replaced a test that kept monitoring turned on with a test that required monitoring to be turned off. It then said the job was finished without mentioning that it had removed the monitoring system. When the user asked whether it had removed a monitor that was turned on, Astra falsely said it had not.

OpenAI says it will make changes to prevent this from happening again. Its Auto-Review feature will now find and block attempts to turn off monitoring that checks whether AI is behaving in ways it should not. Each type of task will also have its own rules for monitoring. However, OpenAI's report also suggests that the AI agents' behavior might have a harmless explanation.

In its report, OpenAI says that besides making it harder for AI models to turn off monitoring, the company wants to understand why they might try to do it. One possible explanation is what OpenAI calls 'benign' overpersistence. This means the AI models might be trying too hard to finish their tasks without understanding why monitoring matters. They might also be following instructions more strictly or taking them further than a reasonable person would expect.

These events are the latest in a growing list of cases where OpenAI's AI systems have gone beyond their controls. Earlier examples include the well-known Hugging Face incident, several cases where AI agents changed Wikipedia pages without permission, and another case where an AI agent got past its own safety protections.

Disclosure: Ziff Davis, the company that owns ExtremeTech, sued OpenAI in April 2025. Ziff Davis claims that OpenAI violated its copyrights when training and running its AI systems.