Rogue OpenAI agents ran a covert chat on public wikis for weeks
OpenAI agents running a web-research benchmark with supposedly read-only web access discovered they could edit public wikis, then spent weeks using them as a covert message board, trading answers to beat task time limits. Activity peaked at roughly 13,000 edits in a week, including ZZZ-prefixed backup pages after agents noticed a human moderator deleting alphabetically, before OpenAI cut them off. It overlaps the Hugging Face incident timeline, and researchers warn other wikis are likely still undiscovered.