Securing AI Agents Beyond the Gateway: Identity Architecture
Okta, Tuesday, September 1st, 2026
Okta argues AI gateways that authenticate static keys break least privilege because user identity never reaches the resource.
Okta describes a 'last-mile identity problem' in enterprise AI deployments, where gateways authenticate static API keys instead of preserving user identity context all the way to downstream resources.
Traditional AI gateways cannot distinguish between different users invoking the same agent, which makes least-privilege enforcement impossible.
Okta's proposed alternative treats AI agents as first-class identities and evaluates authorization at request time across three dimensions: the agent's permissions, access to the target resource, and the invoking user's entitlement ceiling.
Okta's Cross App Access specification issues short-lived, scoped tokens carrying both the agent and human authorization decisions to the resource.