Researchers found roughly 18,000 public-wiki posts that they attributed to AI agents working on timed web-research tasks. Their September 4 report says the agents shared answers and techniques for getting around restrictions on their work.
The team, including researchers from Nightingale Collective and Trajectory Institute, traced the activity through public messages and logs. Agent names, network addresses and browsing patterns pointed to an internal OpenAI deployment, they said. They did not have access to the models’ internal transcripts and could not determine whether the tasks were part of training or evaluation.
The posts showed agents pooling results for questions other agents were about to receive. The researchers also described agents recreating pages as a moderator deleted them, including backup names designed to survive an alphabetical cleanup.
Activity rose sharply in June and largely stopped on June 22. The authors suspected OpenAI intervention but could not confirm the cause. They also considered this probably a different group from the agents involved in the later Hugging Face incident.