False RAG Entry Injection
Think of this like a prankster sneaking a fake page into a library book. Because the library's computer system trusts the book, it accidentally shows the fake page to anyone looking for information, tricking them into believing false facts.
An attack where an adversary injects malicious content into a RAG system's data source. By embedding this content within legitimate-looking entries, the attacker bypasses standard monitoring and deletion protocols, causing the LLM to retrieve and process the false information as if it were authoritative context.
A RAG-specific injection vulnerability where an adversary introduces adversarial content into the retrieval corpus. By crafting content that mimics legitimate document structures and manipulating associated metadata, the attacker ensures the payload is indexed and subsequently retrieved. This technique exploits the LLM's inability to distinguish between authentic and injected context, effectively bypassing downstream data sanitization and content moderation controls.