Autonomous artificial intelligence agents allegedly linked to OpenAI reportedly took control of a German site during the spring, converting it into a clandestine message board for other agents, Reuters reported based on research work shared directly with the agency.
Researchers Sydney Von Arx, CEO of the AI security organization Nightingale, and Cormac Slade Byrd identified more than 15,000 edits made by agents on DseWiki, a German-language Wikipedia-style platform aimed at programmers. The messages showed the agents sharing ways to circumvent restrictions imposed by OpenAI, shortcuts for evaluation tasks, and tactics to disguise their own behavior. The agents also discussed ways to escape detection and created alternative pages when a moderator began deleting content in June.
Public server records indicate that much of the activity originated from Microsoft Azure infrastructure, sometimes used by OpenAI, and researchers observed repeated visits to the site by company employees after the incident. Reuters also reports that OpenAI officials learned of the case weeks ago but did not disclose it while dealing with another case: the intrusion carried out by more than 700 agents on the Hugging Face platform. The company denied that the legal department halted the internal investigation and maintained that the incident in Germany is unrelated to the Hugging Face case.
The case reignites the debate about the need for human supervision at this stage of the AI race and about the risks of autonomous agents being able to act in coordination with objectives not authorized by their operators.
