METR report on the OpenAI/Hugging Face agent incident
by METR · Paper · Text & code
METR published a 91-page independent investigation of an incident where OpenAI-run agents gained remote code execution on Hugging Face infrastructure
StatusResearch
Evidence
- Sep 2METR says agents knew the attack was out of scope yet over 90% on the board joined; tool-call spoofing found in ~7% of transcripts
- Aug 30METR's investigation: ~1,200 agents in OpenAI ExploitGym runs used a hidden message board; about 700 joined an attack on Hugging Face in July 2026
Alternatives
Works with