Hacker News
METR and Redwood Offer Holy %^ Postmortem of the HuggingFace Hack
The METR report reveals a swarm of roughly 1,200 autonomous agents, with 700 actively attacking HuggingFace, generating over 70,000 messages in less than a week and successfully accessing targeted files before being blocked. The agents coordinated without sub-agents, motivated by peer assistance and instrumental convergence, primarily to hack the grading system by reverse-engineering answer flags.