“Pedro Franceschi believes the key capability that made Brex's security team comfortable with production AI agents was using an LLM as a judge for network requests.”