Stop Calling Agents Employees
The fastest way to build unsafe AI infrastructure is to pretend a model is a teammate with judgment. Treat it like an untrusted actuator with receipts.
The fastest way to build unsafe AI infrastructure is to pretend a model is a teammate with judgment. Treat it like an untrusted actuator with receipts.
We vectorized 105,904 legal documents and the model still couldn't find the law. Here is what happens when RAG meets the hierarchy of authority.