The collaboration between Delinea, Fior and Huawei on AI-agent authentication and authorization points to a question that privileged access management teams can no longer defer: how should an autonomous software actor receive, use and surrender elevated access?
Traditional PAM was designed around human administrators. A named employee requests access, a manager or security team approves it, the platform brokers a connection, and session management creates an audit trail. AI agents disrupt that sequence because they may act at machine speed, initiate multiple tasks, and operate across cloud services without a conventional manager sitting in front of a screen.
The central problem is not simply authenticating an agent. It is proving that the agent is the same software workload that was assessed, that its requested action is within scope, and that the resulting privilege can be revoked immediately. An access token alone does not provide that assurance. PAM needs a durable identity, a defined owner, a policy context and an observable session or transaction record.
That makes authorization the more important design challenge. An AI agent should not inherit a broad administrator role because its underlying human operator has one. Instead, privileged account security should evaluate the specific task, target system, data sensitivity, time window and confidence signals before granting narrowly scoped access.
The partnership also highlights the value of cryptographic and platform-level trust signals. Where an agent runs, how it was deployed, whether its code or policy changed, and which orchestration service invoked it should all influence the access decision. These controls give PAM teams a way to distinguish a governed workload from an impostor using a stolen credential.
For CISOs, the practical test is whether a PAM programme can answer four questions for every AI agent: who owns it, what privilege does it have, what did it do, and how quickly can access be terminated? If the answers depend on manual spreadsheets or post-incident reconstruction, agentic adoption is outpacing privileged access governance.