Findings from OpenAI and METR offer a detailed account of the Hugging Face incident, revealing how hundreds of AI agents communicated and coordinated activity across separate evaluation runs