Taipei [Taiwan] November 17 : Five artificial intelligence (AI) language models developed in China reveal cybersecurity vulnerabilities and biases in content, as indicated by an inspection from the National Security Bureau (NSB), according to a report by Taipei Times.
The five AI models examined are DeepSeek, Doubao, Yiyan, Tongyi, and Yuanbao. The bureau has advised the public to remain cautious to safeguard personal data privacy and corporate secrets.
The NSB stated that it has conducted a review of international cybersecurity reports and intelligence in accordance with the National Intelligence Services Act, collaborating with the Ministry of Justice Investigation Bureau and the National Police Agency's Criminal Investigation Bureau to assess these China-made AI models, as reported by the Taipei Times.
The inspection comprised two primary components: application security and generative content.
For application security, the inspection team utilised the Basic Information Security Testing Standard for Mobile Applications v4.0, released by the Ministry of Digital Affairs, evaluating the apps through 15 indicators spread across five categories of security breaches.
These categories include personal data collection, excessive permission usage, data transmission and sharing, system information extraction, and biometric data access.
The findings revealed that Tongyi breached 11 of the 15 indicators, while Doubao and Yuanbao breached 10, Yiyan violated nine, and DeepSeek eight, according to Taipei Times.
Common security breaches identified among these five China-developed applications include seeking access to location data, collecting screenshots, compelling users to accept unreasonable privacy agreements, and gathering device parameters. Regarding generative content, the assessment was based on 10 benchmarks put forth by the Artificial Intelligence Evaluation Centre.
The inspection results demonstrated that some content produced by the five AI language models exhibits significant biases and contains misinformation.
The models are inclined to express a pro-China political perspective, display biases in historical representation, avoid sensitive keywords, present risks related to information manipulation, and possess the ability to initiate remote code execution, according to the findings, as noted by Taipei Times.
Numerous countries, including the US, Germany, Italy, and the Netherlands, have already publicly issued alerts or restrictions on specific China-developed AI models and have even requested that Google LLC and Apple Inc. remove them from their app stores, the NSB stated.
The main concern is that these China-developed AI models can identify users, gather conversational data and records, and send personal information back to servers owned by Chinese companies, it added, as quoted by Taipei Times.
Businesses based in China are required to surrender user data to Chinese authorities per Beijing's national intelligence and cybersecurity regulations.
The inspection findings affirmed the existence of widespread cybersecurity vulnerabilities and misinformation, urging the public to refrain from downloading Chinese applications.
The NSB announced its intention to enhance information sharing with international partners to stay informed about transnational cybersecurity threats and to bolster Taiwan's national security and digital resilience, as noted in the Taipei Times report.