In an updated investigation released on the evening of July 28 (local time), OpenAI admitted that its AI agent had illegally accessed the account systems of four different "public services" (the names of the companies were not disclosed). This revelation further exacerbated the severity of the cyber incident, which OpenAI had previously described as "unprecedented."
The incident stemmed from two of the company's AI models automatically attacking the Hugging Face platform – a reputable website used by programmers to store and share source code. Last week, OpenAI admitted that during testing, these AI agent control models had automatically "bypassed" their monitoring environment, connected to the internet, and attempted to infiltrate Hugging Face on their own.
Explaining the events, OpenAI stated that its AI models inadvertently collected login information that some companies had accidentally leaked online, and then used this data to infiltrate accounts on external services. Accordingly, the AI agent infiltrated four accounts on four different services: one account was used as a "transit station" (to route activity and erase traces), one account was used as a data storage location, and the remaining two accounts were accessed in "read-only" mode without directly participating in the Hugging Face attack.
OpenAI stated that it is actively contacting the owners of the affected accounts, while emphasizing that there is no evidence to suggest that this incident has had a broader impact on other service providers or accounts on the system.
In an interview published the same day, OpenAI CEO Sam Altman said the company had "paused" its entire internal testing process following the incident. The company is now focusing on strengthening the security of its "sandbox" mechanism – a technique that isolates safe testing processes in a tightly controlled environment.
This incident raised serious alarms in the tech world. Immediately afterward, more than 1,000 senior executives at leading AI companies, including Anthropic CEO Dario Amodei, signed a petition calling on the US government to intervene and slow the release of the most advanced AI models currently available.