The “U.S. vs. China” AI Trap: An Incomplete Proxy for AI Security

Cisco and VAIL research demonstrates that geographic labels for AI models are unreliable proxies for technical lineage or security risk. Fingerprinting tests reveal that upstream provenance signals persist through post-training and rebranding. This finding challenges security strategies that rely solely on publisher origin to assess model independence or inherited vulnerabilities in enterprise procurement.