DOES RAG MAKE LLMS LESS SAFE? BLOOMBERG RESEARCH REVEALS HIDDEN DANGERS

Summary from the TLDR InfoSec newsletter

Retrieval Augmented Generation (RAG) can unexpectedly make LLMs less safe by bypassing built-in guardrails. When using RAG, models that normally refuse harmful queries often produce unsafe responses. Generic AI safety frameworks frequently fail to address domain-specific risks in industries such as financial services.