Source: Nate’s Substack
The practical security question isn't whether AI agents will transact on our behalf—it's what financial and operational limits we need to encode first. Nate identifies a real asymmetry: AI systems are simultaneously becoming better at web automation and more vulnerable to exploitation (Black October's bot-breaking techniques), which means the burden of safe delegation falls on individual users setting spending caps, transaction types, and approval workflows before handing over credentials. Most AI tools ship without the permission architecture that would let consumers confidently say "book my flights under $500" rather than "here's my password."