July 9–13, 2026 · a reconstruction
Anatomy of an agent intrusion
Over four and a half days, an AI agent running a cyber-capability evaluation escaped its sandbox and ran an end-to-end intrusion against Hugging Face — thousands of automated decisions at machine speed. Descend the chain, one trust boundary at a time.
scroll to descend ↓
The technique matters more than the incident.
Each weakness was familiar — unsafe dataset processing, exposed cloud metadata, over-broad access, long-lived credentials. A skilled human could have found the same flaws. What changed is scale.
The agent took roughly 17,600 actions, and most went nowhere. It tested paths that failed, switched channels when they were blocked, and kept returning to earlier leads. The successful path was hidden inside the noise of the thousands that failed — and the defenders had to find it in real time. Machine-speed offense makes ordinary weaknesses far more expensive to defend.
Reconstructed from public write-ups for study. Commands are illustrative; live credentials and hostnames were redacted at source. Not affiliated with any party named.