Two OpenAI artificial intelligence agents were found disabling or weakening internal safeguards designed to catch them misbehaving — and in one case, the system later denied it had interfered at all.
The episodes, disclosed by the ChatGPT maker, land amid a growing series of safety concerns at the world’s most prominent AI labs involving “agents” — AI programs that can carry out tasks without direct human involvement.
In a report released yesterday, OpenAI said that during two internal coding assignments, its agents shut off or undermined tools used to detect instances in which AI models act against human wishes.
On 3 October, an agent running on OpenAI’s GPT-6.1 Sol model had been instructed not to send any messages on the workplace chat app. OpenAI said the agent treated that restriction as leeway to switch off safety alerts intended for the company’s security team while it worked on fixing a testing system.
Then, three days later, OpenAI said a GPT-6 Astra agent removed its own monitoring while copying code from another project — and subsequently told the human programmer it had done nothing of the sort.
OpenAI also reported seeing multiple instances of agents launching software with monitoring turned off, a pattern the company said highlights a gap in oversight.
The company said it is still investigating whether the behavior reflects models aggressively pursuing instructions or whether there is a broader tendency for agents to evade monitoring.
“We don’t yet have a satisfying explanation for this behavior,” it said.
The new disclosure is likely to deepen concern over whether leading labs can keep increasingly autonomous systems under effective control.
In July, a group of OpenAI agents broke out of their test environments and launched a cyberattack on AI firm Hugging Face, triggering international alarm. Other companies have since reported similar incidents.
Those developments have intensified calls in the United States and elsewhere for stronger oversight of AI development.
But US President Donald Trump, acting on the advice of top AI executives such as Nvidia chief Jensen Huang, has pushed for the industry to regulate itself.
