Five Steps CISOs Can Take to Secure AI Agents

Summary
A sponsored article recommends securing AI agents through managed identities, tightly scoped access, continuous discovery, intent-based controls, and lifecycle governance.
Key points
- Treat each AI agent as a managed identity with an owner, authentication, defined permissions, and activity logging.
- Scope agent access to the systems, data, and actions needed for its role rather than relying on prompt guardrails alone.
- Continuously discover agent-related tokens, service accounts, and OAuth grants, and map them to the systems they can access.
- Align permissions with an agent’s intended purpose; for example, a ticket-summarizing agent should not need access to export a full customer database.
- Review ownership and access over time, rotate secrets, and decommission agents that are no longer needed.
Article Details
- Defense Focus
- Govern AI-agent identities and constrain their access throughout their lifecycle.
- Detection Methods
- Continuously discover machine and non-human identities, including AI-agent identities.
- Identify agent-related tokens, service accounts, and OAuth grants.
- Map which agents can access which systems and monitor their activity.
- Review whether an agent's permissions remain aligned with its intended purpose.
- Data Sources
- AI-agent identity records and ownership
- API tokens, OAuth grants, service accounts, cloud roles, secrets, and access keys
- Agent activity logs
- Records of agent access to systems and data
- Defensive Actions
- Assign every AI agent a clear owner and authenticate it.
- Explicitly define and tightly scope agent permissions; do not automatically inherit a human's privileges.
- Limit access to the systems, data, and actions required for the agent's stated intent.
- Log and monitor agent activity.
- Continuously review access, rotate secrets, and decommission agents when appropriate.
- Identify and govern shadow AI agents connected to business-critical systems.
