Risks of Using AI Models Developed by Competing Nations

The report warns that the boom in open and offline large language models brings significant risks — including political bias, degraded reliability or insecure outputs from malicious fine‑tuning, hidden backdoors, and supply‑chain attacks (e.g., typosquatting and malicious models on model repositories). It urges organizations to scrutinize model provenance, perform regular security audits and penetration testing on model-generated code, and, where possible, control or distill their own models to manage bias and limit exposure.