Attackers can plant text in documents, tickets, emails, or web pages that a retrieval system later pulls into context, steering answers or triggering tool calls. Vector stores can also leak data across users or tenants when permissions are not carried into the index.
Read in this order
- OfficialLLM08:2025 Vector and Embedding Weaknesses (OWASP GenAI Security Project)Start here: the main weaknesses in vector stores and their mitigations.
- PaperNot what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt InjectionThe paper that showed indirect injection working against real LLM-integrated apps.
- PaperPoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language ModelsMeasured results for poisoning a knowledge base with just a few texts.