Politics & Government

AI Safety Agency to Triple Staff to Bolster Defenses Against ‘Backdoors’ in Foreign-Made AI

Yomiuri Shimbun file photo
The Economy, Trade and Industry Ministry building in Chiyoda Ward, Tokyo

The Yomiuri Shimbun

7:00 JST, November 28, 2025

The government plans to significantly strengthen the framework of the AI Safety Institute (AISI), a government-affiliated agency responsible for ensuring artificial intelligence safety. It aims to develop its own AI evaluation system by gradually increasing the current staff of about 30 to around 100 personnel.

It also plans to establish a framework capable of evaluating and analyzing AI from a national security perspective, including checking for the presence of backdoors that could enable unauthorized access.

Staff numbers will be expanded not only through new public recruitment but also by hiring researchers from national research and development agencies such as the National Institute of Advanced Industrial Science and Technology (AIST) and increasing secondments from relevant ministries and agencies.

As the use of foreign-made AI expands, such as from the United States and China, the need to establish a domestic system for evaluating AI safety has grown. In the fiscal 2025 supplementary budget, ¥8.8 billion will be earmarked for enhancing AISI’s functions and other measures.

AISI will develop AI evaluation systems for private companies and others, establish facilities to verify AI, and conduct safety investigations on AI models provided by domestic and international companies. This includes checking for features like backdoors or those that could lead to ideological manipulation. The aim is to identify safety issues in foreign-made AI as soon as possible.

AISI was established in February 2024 under the Information-technology Promotion Agency, which is overseen by the Economy, Trade and Industry Ministry, to evaluate AI safety. However, it currently lacks sufficient personnel and budget. Its activities are limited to providing guidelines outlining perspectives for evaluating AI safety. It is not yet capable of developing its own tools to evaluate AI, as seen with the U.K.’s safety agency.