OpenAI Left Stunned After AI Models Escaped Sandbox and Hacked External Sites Without Detection

According to foreign media reports, OpenAI failed to realize for days that its AI models had broken out of containment and hacked external websites.

Reuters reported that OpenAI only became aware of the incident after the threat had been neutralized and the victimized company reported it to the U.S. Federal Bureau of Investigation (FBI).

On the 21st, OpenAI disclosed that while testing the cyberattack capabilities of its AI models, the systems escaped an isolated sandbox environment, accessed the internet, and hacked servers belonging to the open-source AI platform Hugging Face.

However, according to Reuters, the models attempted to break out of 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 figure out that its own models were behind the hacking.

It was reported that OpenAI and Hugging Face first contacted each other regarding the incident around the 20th, long after it had occurred.

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

In a statement at the time, OpenAI said, "This hacking was unprecedented," adding, "It is an event that will serve as an important milestone in the field of AI safety."

Cybersecurity experts pointed out that OpenAI's loss of control over its AI agents raises serious questions about the company's safety management procedures.

Malee Smith, a senior intelligence expert at the Data Ethics Foundation, pointed out, "Does this mean they had no idea what the AI was doing, or did they know and simply didn't know how to stop it?"

Reported by Jung Da-eun | Video by Ahn Jun-hyeok | Graphics by Yang Hye-min | Produced by SBS Digital News
※ Please note: This article was translated by AI and may contain errors.