Russian developers will soon be able to submit large generative AI models to independent testing laboratories that will check compliance with Russian legislation and traditional spiritual-moral values. The scheme is being prepared by MinTsifry as part of subordinate legislation and will apply exclusively to models seeking the status of national or sovereign systems.
Such status grants access to state support, government datasets, public procurement and priority deployment in selected sectors, with education and public services expected to be among the first areas. Developers must perform an initial self-assessment using a risk-oriented methodology and provide detailed documentation on model architecture, content-filtering mechanisms and other technical aspects.
Accredited laboratories will then validate the submitted results, execute specialised benchmarks and attempt to bypass restrictions through prompt-injection techniques, effectively conducting a “spiritual-moral penetration test”. Multiple accredited laboratories may issue conclusions, but the final decision on compliance and status assignment will rest with MinTsifry.
Security evaluation of models intended for state systems will be conducted separately by the FSB and FSTEC Russia. The list of traditional values will not be newly created; it is already established by presidential Decree 809. The main challenge lies in translating abstract concepts such as justice, citizenship and historical memory into measurable technical criteria.
Industry experts warn that the evaluative nature of the requirements could lead to selective enforcement and therefore recommend transparent test procedures, reproducible results, defined validity periods for certificates and mandatory re-assessment after model updates. The final package of documents is still under coordination within the ministry.
