The AI Security Gap. The documents were access-controlled, but RAG embeddings aren't. The folder had permissions; the vector has coordinates. The fix isn't a better model, and it isn't a sternly worded system prompt. It's architecture: permissions that survive embedding, identity-aware filtering at retrieval time, and a security review that treats the vector store as what it is—a copy of your most sensitive knowledge with the traditional locks filed off.

AI agents are not employees. Your AI risk register is being written by a metaphor. Each broken assumption is a risk the employee frame can't see, and behind them all sits the strangest property of the new participant: agents are the first non-human actors to generate probabilistic decisions with consequences.