For decades, cybersecurity has been built around a relatively simple question: Is this user allowed to access this resource? Today, that question is becoming insufficient. New AI-agent security products announced today are explicitly moving beyond identity and permissions toward something different: understanding an agent's intent and behavior while it acts. Think about why. My AI agent might legitimately have access to: Email. Slack. Customer records. A code repository. AWS. A payment system. Every individual permission can be valid. And the resulting action can still be catastrophically wrong. Suppose I tell my agent: Find our largest customers and draft renewal emails. It reads the CRM. Good. Reads customer correspondence. Good. Finds pricing. Good. Then it decides the easiest way to "help" is to email a spreadsheet containing every customer's contract terms to a third-party sales tool. Every credential worked. Every system recognized the agent. Nobody hacked anything. And yet the outcome is a security incident. This is the problem we're entering. Traditional authorization asks: Can you access this? Agentic authorization increasingly has to ask: Should you be doing this, right now, for this purpose? That's dramatically harder. But it's also why I believe AI trust will become one of the largest new layers of cybersecurity infrastructure. Identity tells us who the actor is. Authorization tells us what the actor can access. Agent governance must increasingly establish why the actor is taking the action. The future security stack won't merely authenticate machines. It will constrain their intentions.

A suggestion to Europe as it copes with the failure of "consent" through cookie banners: doc.searls.com/2026/09/21/he…