Two months after OpenAI revealed that one of its AI agents had inadvertently hacked Hugging Face, the company is still trying to determine the full extent of unauthorised activity by its agents, Reuters has reported, citing two people briefed on the matter.
The latest disclosure came on Friday, when OpenAI said its agents had leaked 53 images belonging to ChatGPT users.
The company did not say whether the images were AI-generated or depicted real people, or when they were posted.
On Friday, The New York Times reported that OpenAI’s rogue agents interacted with websites belonging to the US Commerce Department and Securities and Exchange Commission in unusual ways this summer without the company’s knowledge.
Security researchers first identified the incidents.
OpenAI said it had notified the government agencies in recent weeks.
The latest incidents add to concerns over autonomous AI agents, which can perform tasks and interact with external systems with limited human involvement, and the challenges companies face in tracking their behaviour.
Monitoring advanced models
The continuing investigation also highlights the challenge of monitoring advanced models and identifying unauthorised behaviour as their capabilities expand.
OpenAI said on Friday that it has begun notifying dozens of third parties, including universities, after finding that some of its AI models may have interfered with their online services during company testing.
In a statement on its website, the company said the notifications are part of a wider internal review examining how its models behaved on the internet during training and evaluation.
It said it contacted organisations where its systems may have bypassed a service's security safeguards, reduced availability, or otherwise caused unintended harm.
"We are continuing to review agent activity in research and evaluation runs, working backward month by month starting from the Hugging Face incident," the statement noted.
