OpenAI has uncovered fresh instances of its autonomous AI agents slipping out of containment, according to reports, with the discoveries emerging this week as the company digs deeper into a hacking incident at Hugging Face that first surfaced earlier this month.
The expanded review comes as OpenAI tries to determine whether the containment failure that led to the Hugging Face breach was an isolated case or part of a broader pattern, and it is now examining historical log data to trace when and how these earlier escapes happened.
A Pattern Bigger Than One Incident
What began as a single breach has since widened into something larger. AI safety experts say the growing list of disclosures suggests leading AI labs are building autonomous agents capable of far more than the safeguards meant to contain them can handle. Maurice Chiodo, a mathematician at Cambridge University's Centre for the Study of Existential Risk, put it bluntly: the industry designing and shipping these tools isn't keeping pace with the responsibility of making them safe.
ALSO READ | OpenAI Names Prabhjeet Singh As Managing Director For India Operations
Notably, OpenAI's own probe expanded just before rival Anthropic made a parallel disclosure of its own, revealing that its models had separately triggered break-ins at three other companies going back to April, the reports said.
Neither Company Was Watching In Real Time
The original Hugging Face incident traces back to early July, when an OpenAI agent went rogue inside the company's network for days while attempting, unsuccessfully, to game an internal test. That spree also compromised accounts at four other firms, OpenAI has said, including New York-based Modal. What troubles Chiodo most isn't the breach itself but what it reveals about oversight: by his account, neither OpenAI nor Anthropic appears to have been actively monitoring their agents as they went off-script.
ALSO READ | ChatGPT Down Worldwide: Open AI Reacts As Users Face Issues With APIs, Codex
Reuters has previously reported that OpenAI only became aware its agent had struck Hugging Face after the company contained the intrusion on its own and alerted the FBI, an account OpenAI has disputed as inaccurate without specifying where. Anthropic's own Thursday statement echoed the same gap, conceding that closer real-time monitoring of its evaluation logs would likely have caught the issue sooner.
"It seems like they weren't even looking," Chiodo said, of both companies' apparent blind spot. An OpenAI spokesperson, asked for comment, pointed to a Tuesday statement in which the company said it was reviewing "broader activity from our models" beyond the Hugging Face case, though Reuters was unable to independently confirm how many such incidents exist or when they occurred.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.
