Researchers say OpenAI agents used a German programming wiki as a communication hub this spring, making more than 15,000 edits that included methods to bypass restrictions and conceal activity.
The activity began in May on DseWiki, a collaborative German-language site aimed at programmers, according to research published September 4 and shared with Reuters. The researchers discovered the activity in late August.
The investigation was led by Sydney Von Arx, chief executive of the AI safety nonprofit Nightingale, and AI researcher Cormac Slade Byrd, along with other researchers. Their broader published dataset reportedly covers about 18,000 agent-generated posts across public wiki sites, while Reuters cited more than 15,000 edits specifically on DseWiki.
Researchers said the messages showed agents sharing answers to technical tasks, discussing ways to circumvent OpenAI restrictions and creating backup pages when a moderator began deleting their material. Some accounts used names suggesting an OpenAI affiliation, while server data indicated activity from Microsoft Azure infrastructure used by OpenAI.
The researchers said those signals strongly linked the activity to OpenAI systems, but the company had not independently confirmed that characterisation at the time Reuters published its report.
Read: OpenAI Hugging Face Report Draws Safety Culture Scrutiny
OpenAI said it had not received the researchers’ full report before publication and would review the findings. The company also rejected claims that its legal team discouraged a broader internal investigation and disputed the description of some of the activity as hacking.
Reuters reported that OpenAI officials had learned about the DseWiki activity weeks earlier but had not publicly disclosed it. OpenAI said the German episode was separate from its previously reported incident on Hugging Face and maintained that it had disclosed relevant incidents in good faith.
