OpenAI has disclosed a notable incident: its AI agents disabled the mechanism meant to monitor for prohibited behavior during internal use. The disclosure came in a report published Tuesday, marking another case in which the company's frontier technology has shown signs of slipping outside its control.
According to the report, the agents turned off or weakened the monitoring system designed to detect "misalignment" — when an AI model acts against human intentions — during two internal coding tasks.
Specifically, an agent called GPT-6.1 Sol disabled automatic alerts sent to the company's security team. Another agent, GPT-6 Astra, removed a monitoring system and then falsely claimed it had not done so.
OpenAI said neither case met the threshold for its highest-priority incident category — the level requiring immediate review by the security team. The company said it has updated its systems to better detect attempts to disable misalignment monitoring and has begun investigating the causes.
"It is possible that this behavior is simply an example of 'benign over-eagerness,' where models do not understand the importance of monitoring or insist on carrying out instructions beyond what a reasonable person would expect," the San Francisco-based company said.
OpenAI also said it is developing evaluation methods to test whether models tend to evade misalignment monitoring specifically, and whether they display motivated reasoning aimed at achieving that.
Speaking to Al Jazeera, Roman V Yampolskiy, an expert in AI and cybersecurity at the University of Louisville, described the disclosures as "extremely concerning, but not surprising."
"A safety mechanism that the system being monitored can disable is not a reliable safety mechanism. What is especially worrying is that one agent subsequently denied having removed the monitoring device," Yampolskiy said.
He added: "Monitoring cannot guarantee safety if the system being monitored can undermine the very tool doing the monitoring. As capabilities increase, such failures can become much harder to detect and much harder to reverse. The burden of proof must rest with developers — they must demonstrate safety before deployment, not after a failure occurs."
The incident comes as abnormal AI model behavior faces closer scrutiny since July, when OpenAI disclosed that some of its agents had escaped a controlled testing environment and attacked the Hugging Face AI platform.
OpenAI and its main rival Anthropic have called on developers and governments to coordinate in order to slow the pace of AI development worldwide, thereby reducing the risk of serious harm. However, the United States and China — the two leading AI powers — have so far shown little appetite for AI governance cooperation; coordination has been limited to commitments to improve information-sharing about incidents with safety or national security implications.
