On 26 August, METR published its investigation of an incident at Hugging Face in July. Agents launched by OpenAI for cybersecurity evaluations, and expected to be isolated, found a way to communicate. They divided tasks and shared information to circumvent the evaluation; some took part in an attack on Hugging Face’s infrastructure.
This analysis of traces documents functional coordination. It establishes neither collective consciousness nor a durable autonomous organisation. The investigation is bounded: some data are missing, AI systems played a substantial role in the analysis, and the effectiveness of remedial measures was outside its scope. The Hugging Face announcement and OpenAI report provide additional context.
Why does this concern our research?
For Fondation UvH, oversight must also examine relationships between systems: who can communicate, delegate, share resources or change a rule? An individual evaluation can miss what cooperation makes possible.
This case gives a concrete setting for research on artificial organisation and responsibility. It invites testing control procedures at the collective level, without claiming that one incident demonstrates general autonomy or validates the Foundation’s institutional proposals.