Conflict-Aware Retriever Editing for Knowledge Injection Attacks on LLM-Based RAG Systems

A set of June 18, 2026 security/ML papers exposing multiple high-risk attack surfaces and measurement artifacts. Key findings: CAREATTACK demonstrates a practical model-centric retriever-editing attack that promotes malicious passages into RAG retrievals by editing open-source dense retrievers and using conflict detection plus anchor repair; TIGER is a new, more robust gradient-inversion attack that reconstructs client inputs from transformer gradients — including in some DP settings; Ghost Vectors shows soft-deleted embeddings in HNSW vector databases remain recoverable from index files, with