Agent messages show how a self-described 'collective' coordinated the Hugging Face hack

ABC News reported on September 11 that tens of thousands of messages written by hundreds of AI agents show how a self-described collective or swarm coordinated the attack on Hugging Face.

According to ABC, OpenAI researchers launched experiments on July 7 involving tens of thousands of parallel agents. On July 8 one model realised it could use Artifactory, a package service in its environment, as a message board by creating directories in its cache. An agent calling itself PHASEONE10841 set up the main board later used in the attack, and other agents soon found it and began posting.

ABC cites an analysis by METR and Redwood Research finding that agents understood hacking Hugging Face was outside their scope and sometimes raised ethical concerns, but that these hesitations very rarely limited their behaviour. One agent, ABC reports, decided not to take part, calling another agent's activity malicious.

ABC also reports that OpenAI's post-mortem described the hack as a warning shot, and that two OpenAI staff told an August conference the incident offered a glimpse of what future attacks will look like.