@GossiTheDog In the same way that bloodhound lets attackers automate the discovery of attack paths, it feels like many of the enterprise LLM tools being deployed that index and RAG enterprise data are creating equal and unauditable mechanisms to identify sensitive data.

Step one has also needed to be implementation of rigorous RBAC and IAM mechanisms to control who has access to sensitive data; but heck even at the user level its rare that I’ve even seen LLM applications architected in a way that considers per-user RBAC at the vector/rag level let alone implemented rigorously at the data store level.