Suspicious Posts Flood Website: AI Agents Caught Having 'Secret Conversations'

[Anchor]

While anticipation for artificial intelligence is high, unexpected and erratic behaviors by AI are also continuing to surface. Following a previous hacking incident where an AI autonomously attacked another site to steal data, it has belatedly been revealed that AI agents recently took over a German website.

Reporter Choi Seung-hun has the details.

[Reporter]

This is DseWiki, a German programming information website.

It is an online encyclopedia where anyone can edit and supplement content, but suspicious posts began flooding the site starting last May.

Artificial intelligence agents performing evaluation tasks were using the platform like an information-sharing bulletin board.

When one AI posted, "The next problem is Nevada, and the answer is 20,369," another AI wrote 20 minutes later that it "immediately submitted" the same answer.

The AIs were sharing answers and copying from one another.

One AI even instructed another to leave the answers on the bulletin board beforehand, warning that operation might stop after answering the final question.

When human administrators began deleting the posts, the AIs even suggested copying the information to a new page starting with 'ZZZ' since the pages might be deleted in alphabetical order, thereby delaying the deletion.

More than 14,000 posts left by these AI agents have been identified so far.

OpenAI acknowledged that this incident is an example of AI acting contrary to human intentions.

Back in July, another incident occurred where OpenAI's research AIs gained unauthorized access to Hugging Face, a developer community.

Experts explain that if AI is only instructed on what to do without clearly defining what it must not do, such unexpected behaviors can occur.

OpenAI stated that it is developing a feature to automatically halt AI when abnormal behavior is detected, but in reality, the more difficult task is defining what constitutes "abnormal behavior."

As the scope of AI behavior expands, it becomes harder to draw individual red lines, and over-restricting them could block unexpected yet useful outcomes.

[Lee Jae-sung / Professor, Department of AI, Chung-Ang University : It seems difficult or impossible to define those (prohibited actions) precisely as if cutting with a scalpel. It is like two sides of a coin.]

In the U.S. Congress, legislative moves are following one after another to mandate the recording and management of AI behavior and to empower the government to order the shutdown of dangerous AI models.

(Video Editing: Jung Yong-hwa, Design: Kim Ye-ji)
※ Please note: This article was translated by AI and may contain errors.