Skip to content
StrikeCyberStrikeCyber
AI Red Teaming

RAG, Model & Training Data Security

Testing the knowledge bases, vector stores and pipelines behind your AI for restricted document leakage, knowledge-base poisoning and model or data extraction.

How it works

Inside RAG, Model & Training Data Security

01

Access Control in Retrieval

Retrieval-augmented generation (RAG) often indexes documents without carrying over who is allowed to read them.

Our methodology

We test whether users can retrieve restricted content through crafted questions, whether tenant and role boundaries hold in the vector store, and whether citations leak sensitive sources.

  • RAG
  • Access control
  • Vector stores
02

Poisoning the Knowledge Base

Anyone who can add content to a knowledge base can try to control what the AI tells everyone else.

Our methodology

We test ingestion paths for knowledge-base and embedding poisoning, planting content that alters answers or carries hidden instructions, mapped to MITRE ATLAS techniques.

  • Data poisoning
  • Ingestion
  • MITRE ATLAS
03

Model and Data Extraction

Fine-tuned models and their training data can be valuable intellectual property, and can contain personal information.

Our methodology

We test for training data and memorized content extraction, model fingerprinting and extraction attempts, and review how model artifacts and datasets are stored and accessed.

  • Model extraction
  • Training data
  • Supply chain
FAQ

RAG, Model & Training Data Security FAQs

What is RAG security?

RAG security is protecting retrieval-augmented generation systems, where an AI answers questions using documents from a knowledge base. The main risks are users retrieving documents they should not see and attackers poisoning the knowledge base to control answers.

What is data poisoning?

Data poisoning is inserting malicious content into the data an AI learns from or retrieves, so it produces attacker-chosen answers or follows hidden instructions. It can target training data, fine-tuning sets or a live knowledge base.

Can attackers extract our training data?

In some cases, yes. Models can memorize and repeat fragments of training data, including personal information. We test whether your models leak memorized content and how your datasets are protected.

What is MITRE ATLAS?

MITRE ATLAS is a knowledge base of adversary tactics and techniques against AI systems, modeled on MITRE ATT&CK. We use it to plan AI attacks and to map findings so your security team can track coverage.

Talk to an operator about your engagement

Scope this with a senior operator. Fixed scope, fixed price, no obligation.

No obligation, no sales pressure. A senior operator replies within one business day.

(877) 657-8496