According to the ENISA Threat Landscape 2025 report, critical sectors subject to the NIS2 directive were the targets of 53.7% of all recorded cyber incidents. In this context, Sovereign AI has evolved from a theoretical concept into an architectural necessity. Architects often make the mistake of equating sovereign AI with a simple local LLM installation, ignoring critical needs for data management, RAG control, and supply chain security.
Why on-premises is only the beginning: architectural challenges of Sovereign AI
Deploying a model "in-circuit" is merely the infrastructure foundation. Since 27.7% of data breaches are related to digital infrastructure, the model must not be a "black box." A sovereign approach requires significant data residency and strict isolation from external networks.
RAG as a threat and opportunity: how to integrate RBAC/RLS into the AI circuit
Retrieval-Augmented Generation (RAG) without integration with existing access rights (RBAC) and row-level security (RLS) creates a risk of unauthorized data leaks. AI architecture must verify a user's access rights to documents directly at the moment of response generation, which prevents the disclosure of confidential information to unauthorized parties.
Supply chain security for models: from verification to isolation
In critical infrastructure, supply chain integrity is of paramount importance to prevent prompt injection or hidden code execution. It is necessary to implement continuous auditing of model inputs and outputs, as well as to use verified components.
NIS2 compliance: how to build an audit-ready system
Compliance with NIS2 requirements when working with AI requires transparency: what data was used for RAG, who had access to it, and how the system responded to queries. To address these challenges, specialists from Intecracy Group—an alliance of independent companies linked by partner agreements and share exchanges—offer architectural design and integration of AI-native components into existing infrastructure. The UnityBase platform, upon which a number of enterprise solutions are built, features a built-in security model with support for RBAC/RLS and supports on-premises deployment, allowing for the integration of AI tools while adhering to corporate policies and regulatory audit requirements.
| Readiness level | Characteristics | Access control |
|---|---|---|
| Level 1: Local deployment | Model runs on-premises, without integration with access rights. | None |
| Level 2: Isolated RAG | Data is contained within the perimeter, access is not controlled at the RLS level. | System |
| Level 3: Sovereign AI | Full integration with RBAC/RLS, query auditing, verified supply chain. | Full (RLS/ACL) |
FAQ
How can NIS2 compliance be ensured when using LLMs?
It is necessary to integrate the LLM into the access management system (RBAC/RLS), ensure infrastructure isolation, and implement continuous auditing of all data operations.
Can existing access rights (RBAC) be integrated with a RAG system?
Yes, this is an architectural requirement for Sovereign AI. RAG must verify a user's rights to access data sources before generating a response.
What are the architectural requirements for on-premises LLM deployment in the public sector?
The primary requirements include full isolation (air-gapped), the use of a unified domain model for access management (RBAC/RLS), and verification of the component supply chain.
Data sources
- ENISA Threat Landscape 2025
- thakicloud.com: Building Sovereign AI for Air-Gapped Public Sector: On-Premises LLM Reference Architecture | Thaki Cloud Tech Blog
- iris.ai: Sovereign AI for Regulated Industries: Keeping Control of Your Data, Models and Knowledge
- thoughtminds.ai: Sovereign AI and On-Premise LLMs - ThoughtMinds
- sysart.com: Sovereign AI: What It Means and How to Build It - SysArt Consulting
- neuraltrust.ai: Sovereign AI Architecture: How to Build | NeuralTrust