Japan's AISI to finish own AI benchmark review by FY2026

AISI was set up in February 2024 as a public institution, and it builds safety standards and evaluation methods while coordinating with relevant ministries and counterpart bodies overseas. Its four strands of work run from creating the benchmark and holding in-house evaluation capability, to gathering and supplying AI-related information, to leading international coordination and raising international competitiveness. Murakami defines the goal as giving Japan what it needs to be the world's most capable country at developing and using AI.

On cyberattacks, Murakami drew a line between what can be shown and what cannot. As of the October 7 interview there was no evidence AI was involved in the attacks, she said, but AI has indeed raised the speed at which vulnerabilities are found, so defenders need AI too — AI does not tire and works around the clock, at a pace beyond humans. AISI cannot cover this on its own, she said, and will cooperate with cybersecurity specialist organizations. Because it is hard to tell from the outside whether AI was used in an incident, incident reporting should follow the existing cybersecurity incident framework, while AISI tracks what attack methods exist and what capabilities AI brings.

Alongside keeping up a system that can keep evaluating many AI models, Murakami pointed to safety assessment at the deployment stage as a focus. The intent is to examine AI agents and physical AI, including the harnesses they run in.