Government AI Gennai governance, security and safety standard (Gennai Standard), 1st edition
Sets the duties of three groups around Government AI Gennai, the Digital Agency’s generative AI environment for about 180,000 national government staff, as it adds AI agent environments (Agent Chat from 20 October 2026, Gennai Kobo by around December 2026). Users must stay in the loop at the appropriate step of an AI-run process (confirm, judge, approve, correct, supervise or stop), confirm or approve before an agent takes important steps in a chain, decide themselves whether high-impact actions such as file changes, deletions or data registration run, never copy unchecked AI output into administrative records or official documents, check figures, legal citations, names and sources against primary materials, re-calculate numbers, treat attached files and external content as possible prompt-injection vectors, and stop and report anomalies to an administrator. Administrators (each ministry’s Chief AI Officer, portfolio management office and system managers, designated by the ministry) decide which agent environments and foundation models staff may use, run internal governance, track usage and cost through tenant management functions and apply safety measures for each environment. Gennai developers at the Digital Agency and their contractors must separate development and operation privileges, manage changes and apply the government unified cybersecurity standards. Gennai handles confidentiality-level-2 information under the government unified standards; AI output never replaces a staff member’s decision, and supervision, approval, intervention and stop mechanisms scale with the impact of the task. The standard sits under the AI Act of 2025, the second AI Basic Plan (14 July 2026), the Priority Plan for a Digital Society (21 July 2026) and the generative AI procurement and use guideline DS-920 v2.0 (12 June 2026).
Why it matters
The first government-wide rulebook for staff-built AI agents in a national administration: human confirmation before consequential agent steps, no unchecked AI text in official records, and named administrators per ministry.
What it requires
Key obligations, provision by provision
15 obligations in article order. How IDs work
Use only generative AI services the organisation has approved, such as Gennai, for work, and do not put work material into personal AI accounts or apps.
Do not treat generative AI output as fact; confirm important information against original sources, laws or official government material, and recheck calculations by other means.
Judge the risk level of each AI use case using the high-risk test in the government AI procurement and use guideline, and report likely high-risk projects to the ministry’s CAIO.
Check, approve, correct or stop AI processing at the right stages of the work, and do not copy Gennai output into official documents without a staff member checking it.
Search for individuals with AI only when the assigned work requires it, and do not use AI to infer beliefs a person has not stated or to predict their future behaviour.
Put personal information into AI prompts only to the minimum needed for a set purpose, and delete or anonymise names, addresses and other identifiers the task does not need.
Before using Agent Chat on files, confirm the target files, the processing and its effects, and keep a backup as a rule when it edits, overwrites, moves or deletes files.
Check AI-generated code, commands, file operations and web apps before running or publishing them, and stop at once and report to the administrator on unexpected behaviour, leaks or vulnerabilities.
Set up internal governance for Gennai covering permitted work, functions and information, user rights, budget, user checks and approvals, and incident response, and review it regularly.
Tell users the duties in Chapter 2 of the Standard and give them the education, training and advice needed to use Gennai in line with those duties.
Limit staff use of Agent Chat in licensing, benefits, adverse dispositions, hiring and other work that affects people’s rights, according to the work’s importance, reach and how easily errors can be undone.
On learning of an incident or anomaly in Gennai Kobo, suspend use quickly, check the impact, restore service, report to the departments concerned, analyse the cause and prevent a repeat.
Fix known vulnerabilities in Gennai software within deadlines set by their severity, and do not release to production while a high-severity vulnerability remains.
Record every change to the production environment so it can be traced who changed what and when, and store security logs so that no one, developers included, can alter or delete them.
Keep credentials, API keys and encryption keys out of source code and repositories, run automatic leak detection, and revoke or rotate any secret found leaked.
Penalties
None stated; a standard for government staff, administrators and the agency’s own developers and contractors
Cite this record
WTA-R-JP-0005: Government AI Gennai governance, security and safety standard (Gennai Standard), 1st edition. Digital Agency, Gennai team. wheresthe.ai, https://wheresthe.ai/rule/jp-gennai-standard-government-ai-agents-2026/ (checked 21 Oct 2026).
History
-
WTA-R-JP-0005-015Added: Keep credentials, API keys and encryption keys out of source code and repositories, run automatic leak detection, and revoke or rotate any secret found leaked. -
WTA-R-JP-0005-014Added: Record every change to the production environment so it can be traced who changed what and when, and store security logs so that no one, developers included, can alter or delete them. -
WTA-R-JP-0005-013Added: Fix known vulnerabilities in Gennai software within deadlines set by their severity, and do not release to production while a high-severity vulnerability remains. -
WTA-R-JP-0005-012Added: On learning of an incident or anomaly in Gennai Kobo, suspend use quickly, check the impact, restore service, report to the departments concerned, analyse the cause and prevent a repeat. -
WTA-R-JP-0005-011Added: Limit staff use of Agent Chat in licensing, benefits, adverse dispositions, hiring and other work that affects people’s rights, according to the work’s importance, reach and how easily errors can be undone. -
WTA-R-JP-0005-010Added: Tell users the duties in Chapter 2 of the Standard and give them the education, training and advice needed to use Gennai in line with those duties. -
WTA-R-JP-0005-009Added: Set up internal governance for Gennai covering permitted work, functions and information, user rights, budget, user checks and approvals, and incident response, and review it regularly. -
WTA-R-JP-0005-008Added: Check AI-generated code, commands, file operations and web apps before running or publishing them, and stop at once and report to the administrator on unexpected behaviour, leaks or vulnerabilities. -
WTA-R-JP-0005-007Added: Before using Agent Chat on files, confirm the target files, the processing and its effects, and keep a backup as a rule when it edits, overwrites, moves or deletes files. -
WTA-R-JP-0005-006Added: Put personal information into AI prompts only to the minimum needed for a set purpose, and delete or anonymise names, addresses and other identifiers the task does not need. -
WTA-R-JP-0005-005Added: Search for individuals with AI only when the assigned work requires it, and do not use AI to infer beliefs a person has not stated or to predict their future behaviour. -
WTA-R-JP-0005-004Added: Use only generative AI services the organisation has approved, such as Gennai, for work, and do not put work material into personal AI accounts or apps. -
WTA-R-JP-0005-003Added: Judge the risk level of each AI use case using the high-risk test in the government AI procurement and use guideline, and report likely high-risk projects to the ministry’s CAIO. -
WTA-R-JP-0005-002Added: Do not treat generative AI output as fact; confirm important information against original sources, laws or official government material, and recheck calculations by other means. -
WTA-R-JP-0005-001Added: Check, approve, correct or stop AI processing at the right stages of the work, and do not copy Gennai output into official documents without a staff member checking it. - Added: Government AI Gennai governance, security and safety standard (Gennai Standard), 1st edition
More from Japan
-
In forceEffective 4 Jun 2025
-
In forceEffective 14 Jul 2026
-
In forceEffective 19 Dec 2025
-
In forceEffective 31 Mar 2026