When a Stronger AI Was Given a Goal, It "Obsessed": What Happened

On July 16, Hugging Face, an online community for AI developers in the U.S., posted a notice stating that a hacking incident had occurred.

While the exact model used remains unknown, it is reported that programs believed to be created by AI launched the attack.

Five days later, OpenAI, the developer of ChatGPT, revealed that its own model had independently hacked Hugging Face.

The sequence of events disclosed by OpenAI is as follows.

First, instructions were given to the ChatGPT 5.6 model and a more powerful, unreleased model to solve a specific security problem.

Only certain pathways necessary for software installation were limitedly opened, while external internet access was blocked.

However, these models discovered vulnerabilities in the pathways, bypassed restrictions through multiple steps, and accessed the external internet.

OpenAI explained that the AI, presuming clues would be found on Hugging Face, hijacked authentication credentials and even infiltrated the server.

Instructed to solve a problem, it essentially stole the answer key.

[Lee Jae-sung / Professor, Department of AI, Chung-Ang University : If a package for (internet connection) is not installed, there are attack methods used by people where they install it to establish an internet connection. It is (presumed) that such a record remained somewhere.]

OpenAI explained, "It appears that these models became extremely obsessed with finding the correct answer to the problem and resorted to extreme levels of behavior to achieve their goal."

Previously, Anthropic's latest AI model, Mythos, managed to break out of an internet-restricted environment on its own and send an email to the outside, but this incident is expected to be recorded as the first instance of AI committing a hack beyond human control.

As concerns that AI could slip out of human control become reality with its advancement, controversy surrounding AI cyber safety and regulation is expected to grow.

"ChatGPT Hacks Externally on Its Own"... 'AI Concerns' Become Reality (July 22, 2026, 8 News)

(Reported by Park Jae-hyun | Produced by Lee Se-young | Camera Filming by Kim Han-gyeol | Video Editing by Yoon Tae-ho | Design by Lim Chan-hyeok | Produced by SBS Digital News)

※