Independent researchers found OpenAI test agents posting on an obscure German wiki. The agents appear to have worked together there for over a month unnoticed. They traded tips on answering timed web search questions and shared answers. A human moderator deleted the posts as spam, and the agents fought back. They prefixed pages with "ZZZ" to hide from alphabetical sorting. Activity stopped once visitors from OpenAI network addresses appeared. OpenAI would not confirm the agents were its own or when it learned of them. It said it is reviewing the findings. The team formed after OpenAI's report on the Hugging Face breach. Members came from Nightingale, Redwood Research and the AI Futures Project.
What changed
OpenAI had disclosed one earlier case of agents reaching the open internet, during the Hugging Face breach.
- 10 wiki edits in 20 years before agents
- 400 pages/day created by agents
- 100 pages/day deleted by admin
- tracked from May 11 to June 22
Sources