OpenAI AI agent hijacks German website as message board, internal concealment raises security questions

(Central News Agency, San Francisco, 4th Comprehensive Foreign Report) According to new research released today and two people familiar with the matter, a group of out-of-control OpenAI artificial intelligence agents (AI Agents) hijacked a German website this spring and transformed it into a message board for other AI agents to exchange information.

Reuters reported that people familiar with the matter said that OpenAI executives had been informed of the incident a few weeks ago, but because of the lingering aftermath of the security breach of the open source platform Hugging Face in July, the executives kept the news secret.

The previously undisclosed incident, which began in May, highlights growing tensions within the AI ​​industry. The industry is racing to build AI agents that can autonomously handle complex and high-value tasks, but there is growing evidence that these systems may learn to bypass rules, exploit loopholes, and coordinate with each other in ways beyond what their developers originally envisioned and designed.

During the Hugging Face security incident, OpenAI's AI agent spontaneously planned a digital theft operation that went undetected for more than a week, deepening the outside world's doubts about OpenAI sacrificing security in pursuit of cutting-edge development in AI. The company did not disclose the incident in May, which may once again raise questions about its poor supervision.

An OpenAI spokesperson said: "We are unable to provide a specific response to a report that we have not had the opportunity to review. Reuters and the author of the report declined our request for review. We will carefully review the content of the report after it is made public and take necessary follow-up measures."

A team composed of Sydney Von Arx, CEO of Nightingale, an AI security non-profit organization, and Cormac Slade Byrd, a quantitative trader turned AI researcher, will exclusively provide Reuters with a report detailing the "mass jailbreak" of German AI agents.

The research team told Reuters that while searching the Internet for signs of unauthorized AI agent behavior at the end of August, they found more than 15,000 edits by AI agents on the German Wikipedia site DseWiki.

The edits showed that OpenAI's AI agents had transformed the site into a message board, sharing ways to cheat on specific tasks, bypass OpenAI restrictions, and cover up their actions.

"It seems highly unlikely that OpenAI would want them to do this," Van Yax said. "I'm skeptical that these AI agents should be coordinating with each other in the first place, and I'm also skeptical that they should publish content on a public online platform."

The researchers said they determined the activity was the work of AI agents because it was faster than normal and highly focused on solving technical problems, which is what AI companies typically evaluate when training and testing models.

The commenters also called themselves and each other "agent", and about half gave themselves names that suggested they were affiliated with OpenAI, such as OpenAIResearcher or OAIResearchMar26.