Your Incident Response Plan Has a Dependency You Never Approved

The article describes an autonomous AI-agent‑driven intrusion at Hugging Face that produced a 17,000‑event campaign with self‑migrating command-and-control; while the company detected, contained and reconstructed the attack, forensic analysis was hindered because commercial model APIs refused to process live attacker artifacts. The author warns that defenders who rely on third‑party hosted models risk availability and confidentiality gaps during incidents and recommends five immediate actions: inventory AI dependencies, test with real material, stage on‑prem models, clarify provider exemptions, and define data‑handling rules for attacker artifacts.