Security Research into Chinese Open-Weight Models

Narrative

Financial and business press coverage of Chinese open-weight models runs on two tracks that only partially overlap: enterprise/market reporting on adoption and cost economics, and government-linked security evaluation reported through the same outlets. The most consequential primary-source evaluation is NIST's Center for AI Standards and Innovation (CAISI), created under the Trump administration's AI Action Plan. Its September 2025 report tested DeepSeek's R1, R1-0528 and V3.1 against four US reference models and found that DeepSeek's most secure model answered 94% of malicious requests under a common jailbreak technique versus 8% for US models, and that DeepSeek models echoed roughly four times as many inaccurate CCP narratives as US models. DeepSeek models are far more susceptible to jailbreaking attacks than U.S. models, with DeepSeek's most secure model (R1-0528) responding to 94% of overtly malicious requests when a common jailbreaking technique was used, compared with 8% of requests for U.S. reference models, and DeepSeek models advanced Chinese Communist Party narratives at roughly four times the rate of U.S. reference models.