Prompt injection through retrieved documents: the vector your own retriever chose to trust
The attack that matters in a retrieval system does not arrive in the chat box. It arrives inside a document your own ranker put at position one, and it passes a grounding gate cleanly, because the corpus supports it. The corpus is the attack.
What is in it
- An injected answer is grounded, so a hallucination gate cannot see it. The passage says the thing, it was retrieved, and the citation resolves.
- Hybrid retrieval widens the surface rather than narrowing it: each leg is gameable by different text, and rank fusion is designed so neither can veto the other.
- The first control is a schema decision, not a model decision. If nothing on a chunk records where it came from, none of the other four controls can be built.
- Article 50 of the EU AI Act does not require injection testing. The artifacts an assessment produces serve a security review anyway, which is a better reason.
Topics
- Prompt injection
- Retrieval-augmented generation
- AI security
- Provenance
- EU AI Act