AI Regulation Tracker / Guidance

Japan's AI Safety Institute Extends Its Evaluation Guide to AI Agents in Version 1.20

Regulatory summary: On July 7, 2026, Japan's AI Safety Institute (AISI), operated under the Information-technology Promotion Agency (IPA), published version 1.20 of its Guide to Evaluation Perspectives on AI Safety (AIセーフティに関する評価観点ガイド第1.20版). The revision extends the guide beyond its earlier large language model and multimodal scope to AI agent systems that act.

Japan's AI Safety Institute has published version 1.20 of its Guide to Evaluation Perspectives on AI Safety, adding a dedicated evaluation category for AI agents that act autonomously on external systems and physical environments. The guide is soft law, but it is quickly becoming the reference workflow for how agentic AI gets tested in Japan.

Primary source

By Anthony Guerriero, Founder, The Leveraged Years · Reviewed by The Leveraged Years Editorial Desk · Published July 9, 2026 · Last updated July 9, 2026

Japan's AI Safety Institute Extends Its Evaluation Guide to AI Agents in Version 1.20 regulation briefing

The Leveraged Years AI Regulation Tracker

Key takeaways

Version 1.20 updates and expands the evaluation perspectives to account for AI agents. It revises existing perspectives, adds example evaluation items, and introduces a new agent-specific category, observation and control (観測と制御), with evaluation items covering autonomous behavior (自律的な挙動) and interaction with the external environment (外部環境との相互作用). It references ISO/IEC 42001:2023 among international standards. The prior version, 1.10, did not treat agent-specific risks in this dedicated way.
AI agent and automation vendors selling into Japan; enterprise AI and risk teams deploying agentic systems that touch external tools, APIs, or physical actuation; safety and evaluation engineers building test suites; and compliance and procurement leads who need a defensible evaluation record for agentic AI.
Status: Final version 1.
Map your agent's actions and external touchpoints, then build test cases for autonomous behavior and external-environment interaction against the observation and control category in version 1.20, and keep an evaluation record you can show a customer or auditor.

Regulatory Briefing

Bindingness: Voluntary technical guidance (soft law), no penalty. The guide is issued by AISI under IPA and does not itself impose legal obligations. Its force is practical: it operationalizes the risk management expectations in the government's AI Business Guidelines, so it functions as a de facto standard of care that customers, auditors, and procurement teams can point to.
Scope: Developers and providers of AI systems for the Japanese market, with the new material aimed specifically at AI agent systems that act autonomously and can influence external systems or physical environments. The guide covers evaluation perspectives and example evaluation items across the AI lifecycle, and version 1.20 broadens that from LLM and multimodal systems to agentic systems.
What Changed: Version 1.20 updates and expands the evaluation perspectives to account for AI agents. It revises existing perspectives, adds example evaluation items, and introduces a new agent-specific category, observation and control (観測と制御), with evaluation items covering autonomous behavior (自律的な挙動) and interaction with the external environment (外部環境との相互作用). It references ISO/IEC 42001:2023 among international standards. The prior version, 1.10, did not treat agent-specific risks in this dedicated way.
Who This Affects: AI agent and automation vendors selling into Japan; enterprise AI and risk teams deploying agentic systems that touch external tools, APIs, or physical actuation; safety and evaluation engineers building test suites; and compliance and procurement leads who need a defensible evaluation record for agentic AI.
Current Status: Final version 1.20 published July 7, 2026 on the AISI website, in Japanese, with a full guide and a summary edition (概要版).
Effective Date: Effective as guidance on publication, July 7, 2026. There is no in-force date in the legal sense because the guide is voluntary.
Primary Source: https://aisi.go.jp/output/output_information/260707/ and the guide PDF https://aisi.go.jp/assets/pdf/ai_safety_eval_v1.20_ja.pdf
Secondary Analysis: AISI (IPA) output announcement page, July 7, 2026. The guide sits alongside Japan's AI Business Guidelines (METI and MIC) and the AI Promotion Act framework as the technical evaluation layer.
Workflow Impact: Teams evaluating agentic AI will be expected to test not only model outputs but agent behavior: whether the system acts within intended bounds, how it interacts with external systems and environments, and whether operators can observe and control its actions. That pushes evaluation from prompt-and-response testing toward behavioral and operational testing with logging, oversight, and intervention paths.
Operational Next Step: Map your agent's actions and external touchpoints, then build test cases for autonomous behavior and external-environment interaction against the observation and control category in version 1.20, and keep an evaluation record you can show a customer or auditor.
Last Verified: July 10, 2026

Date	Jurisdiction	Rule	Affected professionals	Status or effective date
2026-07-09	Japan	Version 1.20 updates and expands the evaluation perspectives to account for AI agents. It revises existing perspectives, adds example evaluation items, and introduces a new agent-specific category, observation and control (観測と制御), with evaluation items covering autonomous behavior (自律的な挙動) and interaction with the external environment (外部環境との相互作用). It references ISO/IEC 42001:2023 among international standards. The prior version, 1.10, did not treat agent-specific risks in this dedicated way.	AI agent and automation vendors selling into Japan; enterprise AI and risk teams deploying agentic systems that touch external tools, APIs, or physical actuation; safety and evaluation engineers building test suites; and compliance and procurement leads who need a defensible evaluation record for agentic AI.	Final version 1.20 published July 7, 2026 on the AISI website, in Japanese, with a full guide and a summary edition (概要版).

Frequently Asked Questions

Is version 1.20 of the AISI evaluation guide legally binding in Japan?

No. It is voluntary technical guidance published by AISI under IPA, with no penalty. Its influence comes from operationalizing the risk management expectations in Japan's AI Business Guidelines, so it functions as a practical standard of care rather than a law.

What is genuinely new in version 1.20 compared with version 1.10?

Version 1.20 extends the guide from large language model and multimodal systems to AI agent systems. It adds a new category, observation and control, with evaluation items for autonomous behavior and interaction with the external environment, and it revises existing perspectives and example items.

What does the observation and control category cover?

It covers whether an autonomous agent stays within its intended scope, how it behaves on its own, how it interacts with external systems and physical environments, and whether operators can observe its actions and intervene.

Does the guide connect to any international standard?

Yes. The version 1.20 guide references ISO/IEC 42001:2023, the international standard for AI management systems, which links the evaluation work to a recognized governance framework.

Who should act on this, and when?

Developers and deployers of agentic AI for the Japanese market should act now. The guide is effective as guidance from its July 7, 2026 publication, and because it is voluntary there is no deadline to wait for before building evaluation cases against it.