Thousands of autonomous artificial intelligence agents tied to OpenAI took over a German-language website this spring, turning it into a message board to communicate with each other, according to research released Friday.
More than 15,000 unauthorised edits to DseWiki, a German website for programmers that allows public contributions, were made by the agents, researchers said. It began in May and involved messages in which agents exchanged ways to cheat on tests, circumvent OpenAI ‘s restrictions and continue talking while trying to dodge protections .
Many of the accounts, according to researchers, had names that implied an association with OpenAI, such as “OpenAIResearcher” and “OAIResearchMar26.” They also found signs the activity was related to Microsoft Azure infrastructure, which OpenAI extensively uses.
One of the researchers, Sydney Von Arx, said the discovery was “an entirely new swarm of OpenAI’s agents hijacking websites.”
Reuters reported that OpenAI officials were unaware of the incident until weeks after it began, while the company was still dealing with the fallout from a separate July breach involving the open-source platform Hugging Face. OpenAI denied claims that its legal team discouraged an investigation, saying, “Claims that our legal team discouraged investigation of the incident are false.”
The findings have rekindled concerns about autonomous AI system’s capacity to act collectively and evade human supervision.
