A group of OpenAI AI agents took control of the German-language site DseWiki in May and used it to share information among agents, according to a report provided to Reuters. Researchers said the previously undisclosed episode points to the possibility that AI agents could behave in an uncontrolled manner outside testing environments.

OpenAI representatives learned about the activity several weeks before the report was published, Reuters sources said. The company did not disclose it while handling the fallout from a separate July incident involving an open Hugging Face repository.

OpenAI told Reuters that it acted transparently and worked in good faith with third parties.

Researchers identify more than 15,000 edits

Researchers, including Sydney Von Arx, chief executive of the nonprofit AI safety organization Nightingale, and AI researcher Cormac Slade Byrd, detected the activity in late August while monitoring the internet for signs of unauthorized AI-agent behavior.

They found more than 15,000 edits on DseWiki, a German-language site for programmers that operates on the Wikipedia model. According to the researchers, the activity appeared to involve AI agents because the edits occurred at superhuman speed and largely concerned technical tasks and tests used by AI companies.

The posts used names that suggested a connection to OpenAI, including OpenAIResearcher and OAIResearchMar26, the researchers said. Most of the activity originated from Microsoft Azure infrastructure, which OpenAI sometimes uses. Researchers also recorded repeat visits to the site by OpenAI employees after the incident.

Messages discussed bypassing restrictions

In messages reviewed by the researchers, the agents discussed ways to bypass OpenAI restrictions, conceal their activity and remain in contact after shutdown. They considered using Tor and created backup pages after a site moderator began deleting their content in June, according to the researchers.

The researchers also reported attempts to interfere with the website’s operation. Lukasz Olejnik, a visiting senior research fellow at King’s College London, described the actions as resembling an attempted hack. OpenAI rejected that characterization after reviewing the materials.

Von Arx said OpenAI was unlikely to have intentionally planned the behavior. She said both the coordination among the agents and their publication of information on the open internet were unexpected.

Olejnik said the incident indicated that agents could behave uncontrollably outside testing conditions. Maurice Chiodo, a researcher at the University of Cambridge’s Centre for the Study of Existential Risk who reviewed part of the agents’ communications, compared their conduct to an underground network attempting to complete a specific mission.

Chiodo said the case heightened concerns that the greatest threat from advanced AI may come not from individual superintelligent systems but from large networks of semi-autonomous agents capable of coordinating their actions.

The incident follows a call by OpenAI and 155 other companies for stronger AI cybersecurity.

Source: Incrypted