Yubico and Delinea linking AI actions to human approval highlights an important PAM design pattern: autonomous systems may execute tasks quickly, but high-impact privilege should remain bounded by explicit policy and accountable human oversight. The approach addresses a central question for CISOs: how can organisations gain the productivity benefits of AI without allowing an agent to become an unreviewed administrator?
The problem is not simply authentication. An AI agent may authenticate successfully and still be authorised to perform the wrong action, against the wrong resource, at the wrong time. Traditional PAM workflows often assume a human requests access, a manager approves it and the session can be recorded. Agentic workflows require richer context: which agent initiated the action, what task it is performing, what data it used and what exact change it wants to make.
Strong authentication, including hardware-backed credentials where appropriate, establishes the agent or operator identity but does not by itself provide least privilege. PAM policy should scope permissions to a task, target and duration. High-risk actions such as changing production configuration, exporting sensitive data or disabling a security control should trigger step-up approval, even when the surrounding workflow is automated.
Human approval must also be designed for speed and quality. If approvers receive vague prompts or too many low-value requests, they will approve reflexively or teams will create bypasses. Approval requests should present the requested action, affected resource, risk context, agent identity, reason and expiry. For repeatable low-risk actions, policy can allow automation while reserving human review for defined thresholds.
Session management provides the evidence layer. AI actions should produce an immutable record of the initiating identity, policy evaluation, approval decision, commands or API calls and final outcome. That record supports incident response, compliance and model-risk investigations. It also makes it possible to detect an agent whose behaviour is drifting from its intended task.
For PAM teams, the Yubico-Delinea pattern reinforces that agentic access should be governed as privileged activity, not treated as ordinary application traffic. Human accountability, short-lived privilege and detailed action telemetry are the controls that make automation safe enough to scale.
Source: SecurityBrief Australia