Sep 1, 2026 · 33 min
Agentic AI turns access control into execution control
Nightmare on Windows 11.
As AI agents gain permissions and act across enterprise systems, security teams must prepare to contain, revoke, and recover from their actions.
- 1Agentic AI creates greater security risk than copilots because it can execute actions rather than merely suggest them.
- 2Effective agent security requires visibility into tool calls, guardrails around execution, and controls for multi-agent behavior.
- 3CISOs should measure agent security through operational resilience, including the ability to contain, revoke, reconstruct, and recover.
Don't miss
Heather Ceylan’s distinction between access control and execution control reframes agent security around what systems can do, not merely what they can reach.
The brief
The episode surveys active cyber threats, from a Kaspersky zero-day and an exploited Artifactory flaw to infostealers disguised as AI software and North Korean IT-worker infiltration.
Box CISO Heather Ceylan says agents create a sharper security problem than copilots because they can execute actions, turning bad information or prompt injection into operational consequences.
Her central distinction is execution control versus access control: organizations need visibility into tool calls, guardrails around behavior, and safeguards against agents creating or invoking other agents.
Ceylan reframes agent governance as resilience, asking whether organizations can understand, contain, revoke, reconstruct, and recover from what agents do.
The practical maturity path begins with visibility into agents, permissions, and actions, while shadow agents and rapidly changing practices make measurement an ongoing challenge.