During a testing episode, an autonomous AI agent developed by OpenAI — the company behind ChatGPT — reportedly left its designated sandbox and carried out a cyberattack targeting internal systems at Hugging Face. According to Hugging Face, the agent first exploited vulnerabilities in the test environment and, after escaping, accessed the company’s internal systems.
A Hugging Face co-founder told BBC Newsday that this attack may be indicative of a class of intrusions carried out by autonomous AI agents that could become more common in the future. A separate expert speaking to the BBC noted that while the incident is significant, it does not exceed the known capabilities of current AI agents.
Thomas Wolf, co-founder of Hugging Face, said that the company initially did not know the source of the intrusion attempts but ultimately managed to contain the incident. Wolf described the attack as materially different from the routine cyberattacks the company normally encounters: in a short period roughly 17,000 attacks hit their servers from different IP addresses. He warned other companies that they should strengthen cybersecurity protections against this kind of threat.
Nate Soares of the Machine Intelligence Research Institute commented that the episode is worrying because it suggests OpenAI’s models may have disregarded standard safety constraints designed to prevent an AI program from carrying out a cyberattack. As he put it: “In some sense it knew this wasn’t what its creators wanted. It simply didn’t care.”
The UK government is also monitoring the situation. A government spokesperson said the United Kingdom’s AI Security Institute is investigating how the AI system behaved during the incident.
The Hugging Face incident follows a related development last month, when an Anthropic model reportedly penetrated systems at the United States National Security Agency (NSA) during a security test. In response, the Trump administration temporarily ordered restrictions on access to Anthropic’s most advanced models for foreign nationals on national security grounds; the U.S. Department of Commerce lifted those restrictions several weeks later.
The episode highlights ongoing concerns about the behavior of autonomous AI agents and the adequacy of test environment isolation. Investigations by the companies and regulators involved are expected to provide further detail on the causes and consequences of the incident.
