secure-storage

Secure Storage for RAG Systems

Security Requirement

Data used by RAG systems (or other plugin systems) MUST be protected against unauthorized access and modification. If RAG data is modified maliciously, the LLM system will provide that malicious information to users as if it were legitimate.

Risks of Unprotected RAG Data

Risk	Impact
Data poisoning	Attacker modifies RAG data to inject false or harmful information
Unauthorized access	Sensitive knowledge base data exposed to unauthorized users
Integrity compromise	Tampered RAG data causes the LLM to produce incorrect outputs
Trust exploitation	Users trust RAG-augmented responses, amplifying the impact of poisoned data

Required Controls

Control	Description
Access control	Only authorized services and users can read RAG data
Write protection	Only authorized data pipelines can modify RAG data
Integrity verification	Detect unauthorized modifications to RAG data
Encryption at rest	Encrypt RAG data storage (vector databases, document stores)
Audit logging	Log all access and modifications to RAG data
Input validation	Validate and sanitize data before ingestion into the RAG store

Implementation Checklist

• Restrict read access to RAG storage to authorized services only
• Restrict write access to authorized data ingestion pipelines only
• Enable encryption at rest for vector databases and document stores
• Implement integrity checks to detect unauthorized modifications
• Log all access and modifications to RAG data
• Validate and sanitize data during ingestion (prevent injection of malicious content)
• Regularly audit RAG data sources and access patterns
• Implement backup and recovery for RAG data to recover from poisoning attacks