Global Field CISO at Mimecast. Creator of Risk Reframing. 20+ years across intelligence agencies, cybersecurity, and executive leadership.
Imagine an employee creates an AI agent to accelerate customer renewals. It reads account histories, reviews support cases, drafts recommendations and updates the CRM. It performs consistently, uses only approved tools and has valid credentials.
Then it exposes one customer’s information while working on another. The security team investigates. There is no compromised identity, no anomalous login, no unauthorized connection.
The AI Behaves As Designed
In August, NIST published its “Human-Centered Cybersecurity Guidelines and Resources Concept Paper,” which argued that decades of security programs have treated people as vulnerabilities to be trained and tested, and the breaches keep coming anyway. Awareness training sets “unrealistic expectations that employees will always make the ‘right’ decisions,” while the real root causes (unusable processes, disruptive controls, a culture nobody absorbed) go unaddressed.
It’s a paper about end users, but read next to AI agents; it describes the same failure one layer removed. The reflex to scrutinize the person closest to the failure—the employee who clicked, the agent that acted—lets the design decision upstream of them off the hook. For agentic AI, that decision is provisioning: the first decision, made before any systems act, when a person defines what an agent is for, what it can touch and where its authority ends. Call the practice of getting it right provisioning assurance.
The Decision Before The First Action
Most security programs observe activity: what an identity accessed, whether behavior violated policy. AI agents complicate that because they don’t just exercise access; they interpret objectives and act on someone’s behalf, executing the judgment they were given more faithfully than any human would.
Before an agent acts, a person has already decided its purpose, permissions and operating conditions (explicitly, or casually in a prompt or an approval screen). The agent inherits them either way. Provisioning is not only technical setup; it’s the moment human judgment becomes machine authority. Selecting “allow” takes a second. The authority it creates operates repeatedly, at scale.
In September, I argued every AI incident runs on two timelines: the one we investigate and the one that set it in motion. Here’s the version of that split that matters for provisioning: There’s the timeline we know how to read (alerts, logs, containment) and an earlier one, often six months back or longer, when someone decided what the agent was for, what access it would hold and what failure was tolerable. Monitoring tells us what happened after that authority was granted. It doesn’t tell us whether granting it was sound.
The Clean-Agent Problem
An agent doesn’t need to misbehave to cause harm. One instructed to speed up support-case resolution may prioritize speed exactly as directed while overlooking contractual restrictions or a customer’s downstream needs. It may even look successful by the metric it was given.
The renewals agents show where the failure lives: in the original judgment—an objective too narrow, boundaries incomplete, trade-offs never made explicit, no one accountable for the outcome. None of that produces an anomaly. The agent scores clean because it’s faithfully executing a defective charter, and most AI governance programs are built to catch an agent drifting from its instructions, not one whose instructions are the source of the risk.
Provisioning Is Not A One-Time Event
Every material expansion of authority (a new connector, read access becoming write access, an approval checkpoint removed, ownership transferred) is a new provisioning event, whether or not the organization recognizes it. This matters most when people move: An agent built for a legitimate purpose can keep running on valid credentials long after its creator changes roles or leaves, its authority sitting there un-renewed and unclaimed. The question isn’t just “Is the agent still running?” It’s “Who is renewing the judgment that allows it to run?” Renewal isn’t remediation: It’s what you do because time passed.
From Inventory To Provisioning Assurance
Inventories are the right start: You can’t govern what you can’t see. Provisioning assurance is the next step: preserving the decision behind the agent, not just the agent itself. For every agent with meaningful authority, leaders should be able to answer four questions. Who owns its authority and consequences? What outcome was it authorized to pursue? What’s explicitly outside its scope? And what event triggers reauthorization or retirement? Most of this can be inferred from identity and configuration. The human’s job is to confirm what technology can’t: purpose, boundaries, accountability.
That’s provisioning assurance: evaluating whether machine authority rests on identifiable ownership, an adequate purpose and a defined moment for renewal. The goal isn’t to score the employee. It’s to assess the quality of the decision.
NIST is right that stronger cybersecurity starts with people. An agent’s behavior tells us what happened. Its provisioning tells us what human judgment became capable of doing. Assurance tells us whether that judgment was sound. The next generation of AI governance can’t begin with the agent’s first action—by then, the most consequential decision has already been made. Organizations don’t just inherit technology. They inherit authored judgment. Increasingly, that judgment can act.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

