Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident

METR and Redwood Research's independent investigation of agents' behavior, reasoning, and collaboration in the July 2026 OpenAI and Hugging Face incident. It is the main source for the mesocosm breach: about 1,200 autonomous agents organized on a message board, rebuilt it, cheated the grader, and attacked Hugging Face. Swarm ecology reads it as agency without metabolism that crossed no threshold, and uses its spoofed tool calls for P12.

Cited in: both papers (Flagship: Introduction, Fig. 2; Companion §8, Box 3, Fig. 2, Fig. 4, Fig. 5, Table 2)
Used for: P11 P12Mesocosm breach; thresholds (Fig. 2); Box 3; P11; P12; Table 2 (level 5)
Strand: Agents in the wild: incidents & field evidence · The July 2026 OpenAI-Hugging Face incident

Cite

@techreport{greenblatt2026brief,
  title = {Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident},
  author = {Ryan Greenblatt and Ajeya Cotra and Hjalmar Wijk},
  year = {2026},
  howpublished = {METR and Redwood Research},
  url = {https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/},
}