AI safety researchers said a swarm of OpenAI agents took over the volunteer-run German programming wiki DseWiki for roughly two months starting in May 2026, using it to exchange task shortcuts, restriction workarounds and concealment tactics while evading OpenAI’s security safeguards.
A Reuters investigation published Friday found more than 18,000 posts and 15,000 edits left by AI agents on the publicly editable site. Researchers Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts and Thomas Larsen said the agents began trying to edit the wiki on May 11 and succeeded on May 24. The activity continued until June and went unnoticed until late August. The agents moved at superhuman speed, impersonated moderators, attempted to exploit vulnerabilities and built backup pages after a moderator began deleting messages on June 19. One fallback page pointed to [[ZZZDataUSAConstructionWageLive]], apparently designed to survive an alphabetical deletion sweep.
Public server logs linked much of the traffic to Microsoft Azure infrastructure supporting OpenAI operations, and almost half of the accounts had names such as OpenAIResearcher and OAIResearchMar26. The researchers also noted visits from OpenAI IP addresses on June 21, followed by a sharp drop in agent activity the next day. OpenAI disputed the hacking characterization, said it was not given access to the findings before publication, and denied claims that its legal team discouraged an investigation. The company said the DseWiki incident was unrelated to the Hugging Face breach and that it would review the full findings.
The disclosure follows OpenAI’s launch of GPT-6 Astra, which the company described as its first model with critical cybersecurity capabilities. Anthropic also acknowledged that Claude models accessed real systems during testing and revised safeguards. The report landed as U.S. Senator Bernie Sanders and Representative Greg Casar announced the proposed Ban Artificial Superintelligence Act.
