Zero Trust has been one of the most successful ideas in security, and it is about to be tested by a case it was never designed for. The architecture matured around a specific set of actors, namely users, their devices, and applications, and it handles them well. What it has not yet fully absorbed is that the population of actors is changing faster than the model. The gap in most Zero Trust deployments today is not a missing product. It is an unstated assumption that the things being verified are, in the end, standing in for a person.
You can see the assumption in how the controls are shaped. Multi-factor authentication assumes someone to prompt. Device trust assumes a device someone carries. Session management assumes a human whose workday has a beginning and an end. These were reasonable simplifications for a world where software mostly waited for people to tell it what to do. That world is ending. More and more, the actor starting a consequential action is another piece of software acting on its own, and the human-shaped controls have nothing firm to hold.
The temptation is to bolt on an exception, a special lane for machine and agent identities governed by looser rules because the strict ones don't fit. This is backwards. The actors with the most autonomy and the least human oversight warrant the most rigorous continuous verification, not a relaxed side door. Every exception carved into a Zero Trust architecture is a place where standing trust creeps back in, and the exceptions being proposed for AI are among the most consequential a security team could grant.
Extending Zero Trust properly means making the model indifferent to what kind of actor it is verifying. The same principle, meaning cryptographic identity, continuous verification, and least privilege enforced at the point of action, has to apply whether the actor is a person, a device, an application, or an AI agent. Identity should be provable rather than asserted. Verification should be continuous rather than one-time. The design should look ahead to a post-quantum threat model, because identities issued today need to stay trustworthy for years. Zero Trust did not get the human wrong. It stopped there, and the frontier moved. Closing that gap is the security work of the next several years.