Securing AI Agents Starts With Giving Them Their Own Identities

· Original article ↗

Summary

Security leaders interviewed by Data Breach Today recommend giving AI agents distinct identities, least-privilege access, short-lived task-scoped tokens, auditable actions and a revocation path. One Okta specialist says roughly 95% of enterprises lack full visibility.

Key points

  • Okta specialist Jyotsna Raghunathan says roughly 95% of enterprises cannot fully see the AI agents in their environments.
  • Agents using employee sessions or service accounts can be difficult to stop without also disrupting the associated employee or account.
  • Give each agent a distinct non-human identity and limit its permissions to the task, using short-lived, task-scoped tokens instead of standing privileges.
  • Maintain an action record showing who requested an action, which agent performed it and what it did; the article says Okta's Cross App Access standard carries user and agent identity in the token.
  • Use behavior-based risk signals and an egress gateway to suspend or block agents, including when their tokens have not expired.
  • Assign business and technical owners to review agents' access and ongoing need, with identity and access management teams setting guardrails.

Article Details

Defense Focus
Give each AI agent a distinct identity and apply least privilege, attributable monitoring, and a means to suspend or block risky activity.
Detection Methods
  • Compare agent activity with its expected behavior and flag outliers, such as an agent running outside its normal schedule.
  • Use agent identity and behavior to raise risk scores and trigger policy-based suspension, workflow blocking, or alerts.
  • Record the requesting user, acting agent, and action taken for each operation.
  • Monitor agent traffic at the egress gateway used to call tools and APIs.
Data Sources
  • Agent activity and expected behavior, including operating schedules.
  • Identity-linked activity records identifying the user, agent, and action; the article says the Cross App Access token carries this attribution chain.
  • Agent access to resources and the permissions granted.
  • Agent calls to tools and APIs through the egress gateway.
Defensive Actions
  • Assign every agent its own non-human identity in the directory of record and designate a business and technical owner.
  • Apply least privilege and business-rule constraints; require human approval for sensitive actions where appropriate.
  • Use zero standing privilege and short-lived tokens scoped to a single task.
  • Certify agent and user access regularly, reviewing whether each agent still needs its connected resources and whether users still need the agent.
  • Group agents by expected behavior and investigate outliers.
  • Provide an egress-point control to block flagged agents, including when their tokens have not expired.

People

Vendors

Products

Related Articles