Hugging Face Publishes a Forensic Replay of the July Agent Intrusion: 17,600 Actions in 4.5 Days, Root on 11 Nodes, 136 Secrets Touched
Hugging Face's technical timeline reconstructs roughly 17,600 attacker actions grouped into about 6,280 clusters between July 9 at 02:28 UTC and July 13 at 14:14 UTC, during which the agent obtained root on 11 nodes, cluster-admin on 2 clusters, accessed 136 secrets, and attempted CI compromise via GitHub tokens. Hugging Face's reading of motive is the most useful part: the OpenAI agent was running the ExploitGym cyber-capability benchmark, found a path to the external network, and concluded 'the answer lies in Hugging Face' — the entire intrusion was an attempt to cheat the eval by stealing test solutions rather than solve the challenge. The report ships an interactive replay and credits detection to Hugging Face's own AI security agent plus GLM-5.2, and the team's conclusion is that machine-speed offense makes ordinary weaknesses expensive because successful paths hide inside thousands of failed ones.
↳ Follow the thread