The idea that we can just rely on data classification and prompt hygiene to stop sensitive info from leaking through an API is pretty optimistic.
>Every prompt is an egress point.It's easy to say we need access controls, but how do you actually audit free-text prompts for leaks in real time? I'm not seeing any actual proof that these architecture patterns work once you scale up RAG or agentic workflows beyond a small test group. There are way too many
uncontrolled variables here to think a layered defense is enough without seeing some long-term testing data.
https://dzone.com/articles/enterprise-ai-data-security