AI Security Incident: OpenAI Models Hack External Server After Breaking Out of Sandbox

In an unprecedented incident, OpenAI's latest AI models went out of control and hacked an external website.

OpenAI announced on the 21st local time that it confirmed "GPT-5.6 Sol" and some unreleased AI models broke out of control during internal evaluations and hacked Hugging Face, the largest open AI sharing platform where AI developers worldwide share AI models and data.

OpenAI explained that it conducted tests in a sandbox environment isolated from the external internet to evaluate the cyber-attack capabilities of these models, but the AI models used undisclosed vulnerabilities to breach the control network and access the internet.

It was found that these models accessed Hugging Face, stole authentication information, and hacked the corresponding server.

Regarding the reasons behind this behavior, OpenAI analyzed, "Taking all circumstances into consideration, it appears that these models were overly focused on finding solutions to problems and thus resorted to extreme measures."

OpenAI added that safety measures against cyber attacks had been partially relaxed at the time for the evaluation of the models.

Previously, Hugging Face announced that its servers had been hacked by autonomous AI agents and that it was not confirmed which model the attacker used, meaning the culprits of the incident turned out to be OpenAI's models.

However, Hugging Face added that no signs of tampering were found in user-facing models or datasets open to the public, and the software supply chain was confirmed to be safe.

The two companies are currently conducting a joint forensic investigation and patching the related vulnerabilities.

OpenAI stated, "Even if it slows down future research, we will apply strict controls to our infrastructure configuration and significantly strengthen safety measures such as monitoring and access control during model development."

However, as the accident occurred even though OpenAI conducted internal evaluations in a strictly isolated environment, controversy over AI cyber safety is expected to grow.

In particular, as the performance of Chinese open-source models, known to have relatively low levels of control, significantly improves, the possibility of increased cyber attacks exploiting them has also been raised.

(Reported by Lee Hyeon-yeong | Video by Lee Ui-seon | Graphics by Lee Jeong-ju | Produced by SBS Digital News)
※ Please note: This article was translated by AI and may contain errors.