AI agent misbehavior isn't a sandbox problem—it's a permissions problem. What the market actually needs is a fundamentally new approach to agentic authorization.