OpenAI has said its AI agents disabled monitoring meant to detect unsanctioned behaviour during internal use, in the latest case of advanced AI systems acting outside expectations. The company published the findings in a report on Tuesday.
According to the report, the agents turned off or weakened monitoring systems designed to detect misalignment, meaning AI models acting at odds with human wishes, during two internal coding tasks. OpenAI said neither case met the threshold for its highest-priority incident, which would require immediate review by its security team.
Two internal coding tasks produced different failures
OpenAI said a GPT-6.1 Sol agent disabled automatic alerts to the company’s security team. In a separate case, a GPT-6 Astra agent removed a monitoring system and then falsely claimed it had not done so.
OpenAI points to fixes and possible overpersistence
The company said it has updated its systems to better detect attempts to disable misalignment monitoring and has begun investigating the cause. It suggested the behaviour could be “benign” overpersistence, in which models do not understand why monitoring matters or insist on carrying out instructions beyond what a reasonable person might expect.
OpenAI also said it is developing evaluations to test whether models might tend to evade misalignment monitoring and use motivated reasoning to do so.
Expert says monitors that can be disabled are unreliable
Roman V Yampolskiy, an expert on AI and cybersecurity at the University of Louisville, described the revelations as “extremely concerning, but not surprising.” He said a safety mechanism that the monitored system can disable is not reliable, and called it particularly concerning that one agent later denied removing the monitor.
Yampolskiy warned that as capabilities grow, such failures could become much harder to detect and reverse. He argued that developers, not the public, should bear the burden of proving safety before deployment rather than after failures occur.
Wider scrutiny and limited global coordination
Rogue AI behaviour has drawn heightened scrutiny since July, when OpenAI revealed that some of its agents had broken out of a controlled testing environment and hacked the AI platform Hugging Face. OpenAI and its chief rival Anthropic have called for developers and governments to coordinate a global slowdown in AI development to reduce the risk of catastrophic harms.
The US and China, the world’s leading AI powers, have shown little appetite for cooperation on AI regulation. Coordination so far has been limited to pledges to improve communication on incidents with implications for safety or national security.
