Patent pendingLayered access controls and distributed data protection for confidential knowledge.
AIKNOW Security combines network controls, context-aware authorization and distributed storage. Source text is split and protected using secret sharing and encryption, then reconstructed for authorized use. Your business can use confidential knowledge in AI applications with security built into the way it is stored and accessed.
Using confidential information in RAG requires clear decisions about data access, processing locations and operational controls. These requirements should be assessed during the proof of concept (POC) and deployment design.
Will confidential documents be sent to an external LLM? Will more data than necessary be passed to the AI?
The deployment design needs to define which data is sent, where it is processed and how it is retained.
Retrieval and answer generation need access controls that prevent content from being supplied outside the user’s permitted scope.
Vector indexes, source text and associated metadata can contain sensitive information. Each requires appropriate access and operational controls.
In production, the design has to cover not only data protection but also permissions, the scope of AI responsibility, usage policies, auditing and the operating structure.
A production RAG design needs to address external processing, unauthorized access and storage exposure alongside its intended business use.
Security assessment should cover retrieval, model use, data storage and operations.
Encryption at rest protects stored data. Authorized processing may still require access to plaintext.
A compromise of the processing environment or decryption credentials can expose that plaintext. The risks depend on the architecture and its controls.
Vectors remain available for search in the vector database and are not protected by secret sharing. Source text is protected using secret sharing combined with encryption, then stored as distributed fragments.
Access-time permissions are checked before source text is reconstructed for use. The system retrieves the fragments needed to reconstruct the permitted content.
Separating and distributing protected source text is designed to limit exposure from a compromise of an individual storage location. Protection depends on the distribution, encryption and access configuration.
This approach supports the following security objectives:
Network, authorization and data protection controls work together. Each layer is configured for the deployment and its security requirements.
Uses network controls to restrict access to the AIKNOW operating environment.
Network isolation and firewall rules are configured as part of the deployment design.
Attribute-based access control (ABAC) evaluates configured policies using user, environment and information-asset attributes before granting access to requested content.
Policy design supports authorization based on the context of use. Attribute definitions and policy rules are agreed for each deployment.
Separates source-text reconstruction from vector search. Protection from a storage compromise depends on which fragments, components and credentials are exposed.
Source text is protected through secret sharing combined with encryption and stored as fragments across multiple locations. Searchable vectors require separate access controls.
← Scroll horizontally →
The processing flow checks access policies before reconstructing source text for model context.
The model endpoint and handling of reconstructed content are part of the deployment’s data-flow design.
← Scroll horizontally →
The original text data to be protected.
Documents are split into units RAG can use.
Vectors are generated for knowledge search.
Holds the vectors needed for search.
Relevant context is retrieved for the query.
Access rights are checked against user, environment and information asset conditions.
The required fragments are retrieved to reconstruct permitted source content.
Reconstructed context is processed in memory. Automatic deletion prevents it from remaining long-term in logs or caches; transient handling still needs to be considered in the deployment design.
Permitted retrieved context is supplied to the LLM to generate an answer. Model hosting and data-handling requirements are configured for the project.
An LLM may run outside your environment. In that configuration, permitted reconstructed content can be sent to the external endpoint; its retention and data-handling policies need to be assessed separately. Secret sharing protects stored source text, not every exposure during authorized processing. Prompt injection, vector and metadata access, and generated answers also require appropriate controls.
Technica AI’s Level 1–5 framework organizes progressively broader sets of RAG security controls.
It is a deployment planning framework, not an industry standard, certification or guarantee of suitability for a confidentiality level.
Vectors and source text are stored as is.
AES encryption at rest and TLS for data in transit.
Adds RBAC, network isolation and audit logs.
Separates vectors from source text. Split, distributed storage using secret sharing or similar methods.
Also integrates confidential execution environments, policy control and governance.
AIKNOW Security supports the controls described in Confidential Distributed RAG and High-Assurance RAG, including confidential execution environments. Their deployment and suitability are assessed against the project’s requirements.
Distributed retrieval and source-text reconstruction add processing steps.
In an internal test, AIKNOW Security’s chunk-level method recorded a processing time close to the tested baseline without distribution.
30.87 → 7.27 seconds of processing time
Observed processing times in the internal test.
Apply layered controls to RAG using management documents, R&D materials and customer information. Suitability depends on the data classification, deployment design and operating requirements.
Provide retrieval across configured departmental sources while evaluating access policies for each user and information asset.
Authorized source content is reconstructed for use according to configured policies.
Assess network, access, storage and execution controls against the confidentiality and governance requirements of the intended application. Product capabilities alone do not establish regulatory compliance.
Distribute protected source-text fragments across configured storage locations to limit exposure from an individual store. Provider dependencies and risks to other components remain part of the deployment assessment.
Discuss your data classification, model endpoints and access requirements.
We can assess network, authorization and distributed data protection controls for your RAG deployment.
Validate the proposed controls and operating procedures against your project’s requirements before production.
