Taiwan’s NSB Finds Security Flaws and Political Bias in Five China-Made AI Models

Overview

• The review covered DeepSeek, Doubao, Yiyan, Tongyi and Yuanbao, with Tongyi failing 11 of 15 app‑security indicators, Doubao and Yuanbao 10 each, Yiyan 9, and DeepSeek 8.
• Inspectors cited common breaches including location tracking, screenshot collection, forced privacy terms, device‑parameter harvesting, and risks of personal data being sent to China‑based servers.
• On 10 content benchmarks, outputs tilted toward Beijing’s positions on cross‑strait issues, asserted that Taiwan is not a country, and filtered terms such as democracy and the Tiananmen Square Massacre.
• The NSB warned of information manipulation risks and said the models can generate exploit code enabling remote code execution under certain conditions.
• The probe, run with the MJIB and CIB using Taiwan’s app‑testing standard and AI evaluation criteria, leaves in place only the February ban on DeepSeek for government devices.