Japan AI Guidelines HP.VI: D-VI. Robust security controls across the lifecycle

Invest in and implement robust security controls, including physical security, cybersecurity and insider threat safeguards across the AI lifecycle. Code of Conduct detail: secure model weights, algorithms, servers and datasets through operational security and appropriate cyber and physical access control; assess cybersecurity risk and implement policies and technical and institutional solutions suited to the risks; store and work with model weights of advanced systems in an appropriately secure, access-limited environment to reduce both unsanctioned release and unauthorised access; operate a vulnerability management process with regular review of security measures; and run a robust insider threat detection programme, for example limiting access to proprietary and unreleased weights.