Securing AI Agents: The Evolution of Identity Security

Identity Security Beyond PAM: Reining in AI Agent Access

This insight originates from the Segura® webinar, “Identity Security Beyond PAM: From Discovery to AI Agents,” featuring Joseph Carson, Evandro Gonçalves, and Bill Willis of IDMWORKS. They unpacked why modern identity security must now encompass humans, service accounts, machine identities, and increasingly, AI agents.

The AI Access Multiplier

AI agents amplify risk because they transform a single human user’s access into multiple, hyper-active digital identities operating 24/7. As Bill Willis articulated, organizations want efficiency. However, if a user with access to financial transfers, sensitive customer data, or production environments spawns an AI agent with those exact privileges, the attack surface expands exponentially. The directive for security teams: Identify the human users whose access poses the highest risk (money movement, production systems, regulated data). Next, map exactly which AI agents or automated scripts can leverage those specific access rights.

The Danger of Privilege Sprawl

Privilege sprawl—leftover access from role changes, overly broad service accounts, and standing privileges—has always been a vulnerability. AI agents weaponize this sprawl by making dormant permissions highly active again. Joseph Carson warned that without a strict adherence to least privilege, an AI agent can spawn dozens of delegated identities utilizing unmanaged, legacy access. As Willis starkly put it: “AI will find all the cracks in your current environment and just blow it apart.”

Essential Remediation Steps:

• Purge obsolete access rights.
• Eliminate standing privileges.
• Enforce separation of duties within applications.
• Vault all privileged credentials.
• Transition to Just-In-Time (JIT) access models.
• Rigorously review all service accounts and machine identities.
• Monitor aggressively for AI agents co-opting human access profiles.

Guardrails Require Hard Access Controls

Prompt engineering and behavioral guardrails are insufficient on their own. Evandro Gonçalves demonstrated this when an AI agent modified and deleted files on his workstation without triggering a single machine-level permission prompt. To be effective, AI guardrails must be backed by hard access controls, approval workflows, and immutable audit logs.

AI Agent Operating Rules

Rule	Application
Don’t Drift	Ensure agents remain strictly within their defined scopes.
Don’t Assume	Never presume an agent’s built-in safety mechanisms are sufficient.
Dig Deep	Thoroughly investigate what underlying access the agent truly utilizes.
Do Not Deploy	Halt deployment if proper monitoring and access controls are absent.
Log Everything	Maintain comprehensive, auditable records of all agent actions.

The Need for Real-Time Identity Signals

Traditional, periodic access reviews are obsolete against AI agents that can spawn, execute a task in milliseconds, and vanish before a quarterly audit occurs. Discovery must transition to real-time event and signal monitoring. Teams must be able to instantly answer:

1. What identity/agent initiated this activity?
2. Which system was accessed, and what data was touched?
3. Was this specific access pre-approved?
4. Do our PAM, IGA, or SOC tools recognize this entity?
5. Can we immediately pause, quarantine, or terminate it?

If real-time discovery is impossible, Gonçalves advises running discovery cycles as frequently as the environment permits—ideally hourly, but no less than daily.

The Human-in-the-Loop Imperative

For highly regulated workflows (GDPR, PCI, HIPAA, financial transactions), full AI autonomy is too risky. Because AI can generate varied outputs from identical prompts, human oversight remains critical.

AI Agent Actions by Control Level

Control Level	Workflow Type
Fully Autonomous	Low-risk, non-sensitive data sorting or reporting.
Human Approval Required	Financial transfers, production changes, accessing regulated PII/PHI.
Strictly Blocked	Actions violating core security policies or exceeding granted privileges.

Auditing Existing AI Agents

If AI is already loose in your environment, abrupt blocking can cripple operations. Begin with a triage audit focusing on agents with the highest potential business impact.

Triage Checklist:

• Inventory: Exactly which agents are currently active?
• Ownership: Who is the designated human owner for each agent?
• Reach: What systems and networks can these agents touch?
• Risk: Which agents possess the capability to significantly disrupt business operations?

Connecting the Security Ecosystem

Managing AI risk requires cross-platform intelligence. Identity security beyond PAM means your tools must talk to each other: PAM (privileged access/vaulting), IGA (lifecycle/ownership), Access Management (MFA/SSO), the SOC (behavioral alerts), and DevSecOps (pipelines/secrets). A unified approach ensures you can always answer who launched an agent, what it changed, and if it was authorized to do so.