Policy & standardsThu, 19 Nov 2026
Japan expands AI safety evaluation to agents: what autonomy means for security
Japan AISI’s July guide adds agent-specific evaluation perspectives. November policy discussions and industry events highlight both AI-enabled attacks and the security of AI systems.
Article
From model evaluation to agent behavior
On July 7, 2026, Japan’s AI Safety Institute (AISI) published version 1.20 of its Guide to Evaluation Perspectives on AI Safety. Responding to the spread of AI agents, the revision adds perspectives on autonomous behavior and interactions with external environments, under a focus on observation and control. It offers evaluation guidance for developers and providers; it is not a newly enacted law or mandatory certification.
November: attack speed and defensive capacity
Japan’s second ministerial cybersecurity meeting on November 6 adopted an emergency measures package addressing faster attacks in the AI era, basic safeguards, information analysis and response capability. Tokyo’s cybersecurity portal summarizes the package. Separately, the Frontria symposium scheduled for November 26 lists AI safety evaluation, AI-assisted attacks and AI-enabled defense; its program plans a JAI-Trust presentation by Japan AISI’s deputy director. The agenda describes a future event, not remarks already delivered.
What buyers and deployment teams should test
When an agent can call external tools, access data or act on a person’s behalf, purchasing and engineering teams should map its tools, identities, data access and network boundaries. Test whether unexpected actions can be logged, paused, reviewed by a person or stopped. Revisit the assessment when systems change, and make supplier responsibilities explicit. These are TWISA’s practical recommendations informed by the guide, not purchasing requirements prescribed by AISI.
Opportunities for Taiwan–Japan collaboration
AI safety assessments, agent red teaming, identity and tool authorization, behavior monitoring and investigation records can be useful starting points for security providers discussing services with Japanese organizations. Actual opportunities depend on customer needs, partnership terms and formal procurement notices; the guide itself is neither a tender nor proof of market access. Teams can prepare bilingual test scenarios, responsibility maps and verifiable outcomes for collaboration.