AI agents are the new privileged users. The identity stack still treats them as service accounts.

Every new class of digital worker arrives faster than the identity stack that has to govern it — and AI agents are the step-up that makes that lag indefensible. The shape is not new; the speed is.

The mechanism runs the same way every time. Classify by blast radius — what the agent touches, and the worst case if it is hijacked. Scope credentials to that radius: least privilege, time-boxed, task-specific. Audit it like a person — who provisioned it, what it is doing, what it has read. Sunset it like a contractor — a defined off-ramp, rotated secrets, a named owner.

What breaks is not any one stage. It is the absence of ownership across them. Agents get spun up in a customer service or development workflow, inherit the credentials of the user who deployed them, and nobody in the IAM team is on the hook for retiring them. Palo Alto Networks' 2026 Identity Security Landscape Report puts a number on the lag: machine identities already outnumber human identities 109 to 1 in Australia, AI agent identities are projected to grow 85% over the next year, and roughly 40% of those agents already hold access to organisational data. Nine in ten organisations reported an identity-related breach in the last twelve months.

The takeaway is one question for the next IAM review: is there a named owner for every non-human identity, and does that owner run the same lifecycle as for a human hire? If not, the next breach will not look like a hack. It will look like an account that was never supposed to still exist.