Frontier-Model Distillation Becomes a Government-Flagged Security Threat

A September 8 joint NSA, CISA, and FBI advisory alleges that China-based AI firms systematically distilled capabilities from major U.S. frontier models using fraudulent accounts, APIs, cloud services, aggregators, proxy infrastructure, prompt injections, and agentic workloads. Anthropic now alleges nearly 200 million exchanges, chain-of-thought extraction, and live routing of Chinese user traffic through Claude, but the scale, attribution, and privacy implications remain insufficiently independently corroborated.