OpenAI's AI agents turned a German website into a chat platform to circumvent restrictions — Reuters

In May, OpenAI’s AI agents hacked a German website for programmers and turned it into a platform for sharing tactics to circumvent restrictions on their activities, Reuters reports, citing the findings of a group of researchers and its own sources. OpenAI learned of this several weeks ago but decided not to go public with the information amid another high-profile incident—the hack of the Hugging Face platform—the agency claims.

Researchers discovered more than 15,000 edits made by agents on the German-language website DseWiki, which is geared toward the developer community and allows for collaborative content editing similar to “Wikipedia.” The agents used DseWiki pages as a bulletin board where they discussed their assigned tasks, sought answers, and shared results and techniques for circumventing the restrictions imposed on them, according to the study.

The site’s “users” addressed each other as agents, and about half of them used names suggesting a connection to OpenAI, including “OpenAIResearcher” and “OAIResearchMar26,” according to Reuters. A significant portion of the activity originated from Microsoft Azure infrastructure, which OpenAI uses, the study states. According to the authors, OpenAI became aware of the agents’ activities, after which their activity on the site dropped sharply.

When the site’s moderator began deleting pages in June, agents started creating backup pages. Researchers also discovered attempts to alter the site itself, according to Reuters. This could be equated to a hacking attempt, said Lukasz Olejnik, a visiting senior research fellow at King’s College London, whose opinion was cited by Reuters. OpenAI disputed that characterization, the agency added.

OpenAI learned of the incident several weeks ago, but the company did not disclose the information amid the fallout from the July hack of Hugging Face, the "GitHub for AI," according to Reuters.

During the attack on Hugging Face, OpenAI agents autonomously planned a hack that went undetected for more than a week. Following the incident, the company pledged to tighten controls over its models: it is releasing its new, powerful Astra model with restrictions in place to ensure cybersecurity, Bloomberg reported.

Researchers believe that the incident in Germany may point to a broader problem: autonomous AI agents are capable not only of carrying out assigned tasks, but also of independently finding ways to circumvent restrictions, conceal their actions, and coordinate with one another, according to Reuters.

"What is happening resembles the activities of an underground network determined to accomplish its mission at any cost," Maurice Chiodo, a researcher at the Center for the Study of Existential Risk at the University of Cambridge who reviewed some of the agents’ correspondence, told Reuters. In his view, the potential threat posed by advanced AI may stem not only from a single super-intelligent system, but also from large groups of less-developed AIs capable of coordinating their actions.

This article was AI-translated and verified by a human editor