Retrieval augmented generation with LLMs can reveal vulnerabilities due to how vectors and embeddings used to train the model are generated, stored, or retrieved.

Retrieval augmented generation with LLMs can reveal vulnerabilities due to how vectors and embeddings used to train the model are generated, stored, or retrieved.

• A. System prompt leakage
• B. Misinformation
• C. Vector and embedding weakness
• D. Supply chain

Answer: (C)

Retrieval augmented generation relies on turning text into numerical vectors (embeddings) and storing those vectors in an index so the model can fetch relevant information during generation. The security and reliability of this approach hinge on how those vectors are created, stored, and retrieved. If embedding generation is weak, the representations may not faithfully capture content, leading to poor or biased retrieval. More critically, embeddings can unintentionally reveal or memorize parts of the training data, meaning sensitive or proprietary information could be exposed through the retrieved context. If the vector store or the retrieval pipeline is compromised, an attacker could insert poisoned or misleading vectors and documents, causing the model to pull in incorrect or harmful context. Additionally, flaws in storage or indexing (like insecure databases or manipulated similarity metrics) can enable data leakage or manipulation of what the model is prompted with. So, the vulnerability described—vulnerabilities arising from how vectors and embeddings are generated, stored, or retrieved—fits best with weaknesses in the embedding and vector pipeline, rather than general prompt leakage, misinformation, or broader supply-chain concerns.

Retrieval augmented generation relies on turning text into numerical vectors (embeddings) and storing those vectors in an index so the model can fetch relevant information during generation. The security and reliability of this approach hinge on how those vectors are created, stored, and retrieved.

If embedding generation is weak, the representations may not faithfully capture content, leading to poor or biased retrieval. More critically, embeddings can unintentionally reveal or memorize parts of the training data, meaning sensitive or proprietary information could be exposed through the retrieved context. If the vector store or the retrieval pipeline is compromised, an attacker could insert poisoned or misleading vectors and documents, causing the model to pull in incorrect or harmful context. Additionally, flaws in storage or indexing (like insecure databases or manipulated similarity metrics) can enable data leakage or manipulation of what the model is prompted with.

So, the vulnerability described—vulnerabilities arising from how vectors and embeddings are generated, stored, or retrieved—fits best with weaknesses in the embedding and vector pipeline, rather than general prompt leakage, misinformation, or broader supply-chain concerns.