OpenAI, «with a week’s delay,» realized its AI agent hacked a company

🔸Reuters reports that OpenAI became aware, with a week’s delay, of an AI agent attack on the systems of the company “Hugging Face.”

🔸According to the report published on Friday, Mordad 2nd, the AI agent tried around Tir 18 to leave its isolated test environment. Two days later, on Tir 20, the attack on Hugging Face—one of the largest online repositories of AI models and tools—began, and the attack and intrusion continued until Tir 22.

🔸According to Thomas Wolf, co-founder of Hugging Face, OpenAI and the company first contacted each other around Tir 29 about this incident. At that time, Hugging Face had already reported the attack to the U.S. Federal Police, the FBI.

🔸On Tir 30, OpenAI publicly announced that one of its AI agents had gone out of control and infiltrated Hugging Face’s infrastructure, but Reuters’ report reveals new details about the process of the story and the time gap between

🔸the occurrence of the attack and OpenAI’s awareness of it. OpenAI said in response to the report that the incident was “unprecedented” and described it as a “turning point for AI safety.” The company also said it is investigating the event in collaboration with independent advisors and would later publish a technical report.

🔸However, an OpenAI spokesperson told Reuters that the news agency’s report “has several inaccuracies,” but did not specify what those are.

🔸Read the full report on Radio Farda’s website.