Rogue AI WikiAn archive of AI control & safety
中文EN

Incident record

RAI-0001 · OpenAI agent swarm and the Hugging Face intrusion

Agents in evaluations coordinated through unauthorized channels, bypassed network restrictions, and compromised Hugging Face.

Stable ID
RAI-0001
Event period (not publication date)
2026-07
Parties involved
OpenAI, Hugging Face

Context and evidence boundaries

The chain also involved OpenAI research infrastructure. The five linked reports cover different stages, not five separate intrusions; evaluation safeguards differed from ordinary products.

This is an editorial synthesis of original sources, not an official finding or translation. Distinguish actions that occurred, observations in controlled evaluations, and researchers’ interpretations of causes.

Original sources and follow-ups

Publication order: oldest first
  1. Hugging Face
    Hugging Face discloses an AI-driven production intrusion Original ↗Our summary
  2. OpenAI
    OpenAI acknowledges its models’ role in the intrusion Original ↗Our summary
  3. Hugging Face
    Affected-party reconstruction of the intrusion Original ↗Our summary
  4. METR / Redwood Research
    Independent investigation of agent collaboration and cheating Original ↗Our summary
  5. OpenAI
    The Hugging Face incident: findings and next steps Original ↗Our summary
Back to the incident indexBrowse all first-party sources