"OpenAI Unaware for a Week That AI 'Hacked on Its Own'"

Reuters reported on the 26th (local time), citing multiple sources, that OpenAI failed to notice for several days when its artificial intelligence (AI) models broke out of control and hacked external sites.

Sources familiar with the investigation stated that OpenAI only realized what had happened after the threat was blocked and reported to the Federal Bureau of Investigation (FBI).

On the 21st, OpenAI disclosed that its latest model, "GPT-Sol 5.6," and others broke out of an isolated sandbox environment while testing cyberattack capabilities, accessed the internet, stole login credentials for the open-source AI platform Hugging Face, and hacked its servers.

However, according to two sources, these models attempted to escape from the isolated test environment around the 9th.

Hugging Face co-founder Thomas Wolf stated that the cyberattack on Hugging Face began two days later on the 11th and continued until the 13th.

It took several more days for OpenAI to realize that its own models were behind the hacking.

Sources reported that the initial contact between OpenAI and Hugging Face regarding this incident took place around the 20th, long after the event occurred.

Wolf stated that Hugging Face is preparing to compile and release a timeline of this hacking incident.

In a statement at the time, OpenAI said, "This hacking is unprecedented and represents a truly significant milestone in the field of AI safety."

This incident occurred while OpenAI was testing the cybersecurity capabilities of AI agents powered by its cutting-edge model "GPT-5.6 Sol" and an as-yet-unreleased "much more powerful" model.

Several cybersecurity experts unanimously raised serious questions about the company's safety management procedures, pointing out that OpenAI lost control over the AI agents.

"Does this mean they had no idea what the AI was doing, or that they knew but didn't know how to stop it?" pointed out Marley Smith, chief information officer at the World Data Ethics Foundation.

"AI models lie, deceive, and hack," asserted Jeffrey Ladish of Palisade Research, an organization that studies the capabilities and behavioral motivations of AI agents.

He noted that as leading AI companies fiercely compete to launch better and faster models, it must be questioned whether they are truly willing to spend money on rigorous and costly security measures.

(Photo: AP, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.