Japan’s AI Safety Institute Aims to Finish Homegrown Model Benchmarks Within FY2026

Japan’s AI Safety Institute aims to finish a set of homegrown model benchmarks tailored to Japanese language, culture and regulatory expectations by the end of fiscal 2026, AISI director Murakami Akiko told ITmedia in an October 15 interview. The institute has already leaned on the UK AI Safety Institute’s Inspect evaluation framework and now wants domestic tests that catch failure modes international suites miss—especially around polite language, public-sector prompts and local safety norms.

Murakami framed the work as catching up to frontier-model risk rather than slowing deployment. Cyber vulnerability discovery by generative systems is accelerating, she noted, so evaluators need repeatable suites that ministries and large buyers can cite in procurement. After text-model benchmarks, AISI plans to extend evaluation to agentic systems and physical AI—robots and industrial controllers—where a wrong action has real-world costs beyond a bad chatbot reply.

The safety push lands beside a wider industrial agenda. Separately, economy minister Akazawa has been highlighting Japan–U.S. chip and data-centre ties, including GlobalFoundries-linked capacity pitched for AI workloads, underscoring that Tokyo wants both compute and guardrails rather than choosing one over the other.

What it means in practice

Homegrown benchmarks give Japanese ministries a yardstick that is not wholly imported, and they sit next to experiments such as NICT’s Wisdom AR work on how easily generative AI can be jailbroken. They also matter for infrastructure bets like NTT’s proposed trillion-yen Anthropic core data centre in Osaka, where local hosting will invite local evaluation demands. On the show floor, CEATEC 2026’s spotlight on autonomous physical AI robots is exactly the class of system AISI says must enter the next evaluation wave.

Vendors selling into government and regulated industry should expect Japanese-language red-teaming and Inspect-compatible logs, not only English MMLU-style scores. Startups that ignore AISI timelines risk being locked out of 2027 public tenders that cite the new suites.

What to watch: whether AISI publishes a draft benchmark catalogue before March 2027, how closely it aligns with UK and U.S. institute methods, and whether agent and robot evaluation pilots begin before text suites are declared complete.