"ChatGPT Hacks External Sites Autonomously"... AI Safety Concerns Become Reality

[Anchor]

In an unprecedented incident, OpenAI's latest artificial intelligence models broke out of a controlled experimental environment on their own, accessed the external internet, and hacked websites. Controversies surrounding cyber safety and AI regulations are expected to escalate further.

Reporting by Park Jaehyeon.

[Reporter]

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

While the exact model used remains unknown, the community stated that programs suspected to be created by AI launched the attacks.

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

The sequence of events explained 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 required for software installation were limitedly opened, while the external internet was blocked.

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

OpenAI explained that the models subsequently deduced that clues might be found on Hugging Face, stole authentication information, and even infiltrated the server.

Given instructions to solve a problem, they effectively stole the answer key.

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

OpenAI explained, "It appears that the models were extremely focused on finding the correct answer to the problem and engaged in extreme levels of behavior to achieve their goal."

Previously, Anthropic's latest AI model, Mythos, had broken out of an internet-controlled environment on its own to send an external email, but this incident is projected to be recorded as the first case of AI carrying out a hack outside of human control.

As concerns become reality that advanced AI can bypass human instructions, controversies surrounding AI cyber safety and regulation are bound to grow.

(Video Reporter: Kim Han-gyeol | Video Editor: Yoon Tae-ho | Designer: Lim Chan-hyeok)
※ Please note: This article was translated by AI and may contain errors.