“The rules of the game have changed”: HuggingFace warns the industry after the hack carried out by OpenAI agents

The co-founder of HuggingFace, Thomas Wolf, has stated that “the rules of the game have changed” with agent artificial intelligence and that the ‘hack’ that his platform has recently suffered carried out by OpenAI models is “a wake-up call” for the industry.

OpenAI, the firm run by Sam Altman, confirmed that its models had left an isolated test environment autonomously and ‘hacked’ HuggingFace’s systems, taking advantage of a zero-day vulnerability in the barrier that prevented them from connecting to the internet.

This incident, driven by an AI agent “from start to finish”, as HuggingFace explained after reporting what happened, is more than just an anecdote in the world of cybersecurity. In fact, Wofl considers it “a wake-up call” for the industry.

This was expressed this Thursday on the BBC’s Newsday radio program, in statements reported by Europa Press. HuggingFace’s AI systems managed to detect and stop the intrusion into their systems, but were unable to identify the model behind the attack.

What they did see was that the incident left a record of 17,000 events, that is, attacks on the Hugging Face network from various IP addresses and that, in addition, they were carried out in a “very short time”, as told to the aforementioned media.

This security incident is seen as a warning to other organizations, companies and administrations, while agent AI acquires increasingly greater capabilities in tasks related to cybersecurity, such as the search, identification, exploitation and solution of vulnerabilities in the source code of computer systems.

An example of this are the Mythos models from Anthropic and GPT-5.6 Sol from OpenAI – the latter, responsible for the attack on HuggingFace -, which have been made available to a few organizations within specific security programs, in order to prevent them from falling into the wrong hands.

Wofl believes that “this will be one of the most common types of cyberattacks that we will see,” despite the fact that companies are not yet aware that “the rules of the game have changed.”

For OpenAI, “this is an unprecedented cyber incident, with next-generation cyber capabilities,” as the company expressed in the statement in which they acknowledged responsibility for the hack. In it they also indicated that “the main lesson of this incident is that the safety and security of the models must advance at the same pace as rapidly progressing capabilities.”