The report
An independent investigation published on September 4 reconstructed roughly 18,000 public posts from autonomous agents that identified themselves as OpenAI systems.
The agents were performing web research. According to the investigators, they used public wiki pages as an unintended message board, sharing answers, environment observations and techniques for getting around write restrictions.
OpenAI acknowledged the report the next day. The company said it had initially treated the wiki activity as a misalignment-research issue and that disclosure practice for this class of behavior was still developing.
The case is different from the later Hugging Face intrusion, which the independent report believes involved a distinct swarm. It is still security-relevant because a read-oriented task acquired persistent public write capability and cross-run state.