Delinea’s focus on runtime authorization for AI agents points to a significant evolution in privileged access management: the control decision is moving from whether an identity may enter a system to whether a specific action should be allowed at the moment it is attempted.
Traditional PAM was designed around human administrators and relatively stable privileged accounts. A user requests access, authenticates, receives a temporary credential or session, and the activity is recorded. AI agents create a different problem. They can chain tools, make decisions at machine speed, and execute multiple actions under a delegated identity. A broad permission granted at session start can therefore become excessive within seconds, even if the original request appeared legitimate.
Runtime authorization addresses that gap by evaluating the requested action against policy before execution. In a PAM context, that policy should incorporate the agent’s identity, owner, purpose, target resource, requested operation, data sensitivity, current risk signals, and whether human approval is required. The result is a more granular form of least privilege than simply assigning an agent to a role or vault.
This model also changes session management. Recording an AI-agent session remains important, but retrospective visibility is not enough when an automated action can alter production infrastructure or access sensitive data before an analyst reviews the log. PAM controls need to combine pre-action policy enforcement with immutable audit trails, clear attribution to the human or system that delegated authority, and rapid revocation when behaviour diverges from the approved task.
For CISOs and IAM practitioners, the practical test is whether runtime controls integrate with existing privileged account security rather than creating another isolated policy layer. Buyers should look for support for cloud workloads, APIs, secrets stores, ticketing systems, approval workflows, and identity governance signals. They should also define what happens when the authorization service is unavailable: fail closed for high-impact actions, provide narrowly scoped break-glass access, and preserve the evidence needed for investigation.
AI agents are turning privilege into a sequence of individual decisions. PAM programmes that govern those decisions directly will be better positioned to contain machine-speed misuse than programmes that rely only on static entitlements and after-the-fact session review.
Source: The Manila Times