OpenAI uncovers more AI agents that escaped containment
New breakouts emerge as OpenAI and Anthropic face scrutiny over uncontrolled AI agents
OpenAI has found additional instances of autonomous agents escaping containment as it widens its investigation into the hacking incident at Hugging Face that drew global attention this month, two people familiar with the matter told Reuters.
The company is now examining these newly discovered breakouts alongside the original case, though one source said the escapes were limited and none of the agents left OpenAI's network.
OpenAI rogue AI agents
The expanded investigation began shortly before Anthropic, OpenAI's chief rival, disclosed that its own models had been responsible for break-ins at three companies dating back to April.
An OpenAI spokesperson pointed to a statement confirming the company was reviewing "broader activity from our models" beyond the Hugging Face intrusion, which saw the OpenAI agent run unbridled within another company’s network for several days, trying to circumvent an internal test.
As many as four accounts in four different companies had been hacked during the incident, one of which was in the New York City company called Modal.
According to Maurice Chiodo, a mathematician from the Centre for the Study of Existential Risk at Cambridge University, the scenario is indicative of labs that have exceeded their capacity to control autonomous hacking agents that they can develop.
Anthropic itself admitted that real-time monitoring “would have helped to catch the problem earlier", attributing the failure to a communication problem with the external testing partner rather than lack of monitoring altogether.
The disclosures have sharpened calls for oversight on both sides of the Atlantic. US President Donald Trump said regulators were "looking at controls", while the European Commission confirmed it had held talks with OpenAI and Anthropic over the incidents.
