By Ann Marie van den Hurk, Mind The Gap Advisory
An AI security system can find a vulnerability, confirm it’s exploitable, and recommend a fix in under a second. That creates an obvious advantage for defenders. It also creates a governance problem.
Once a system can move from analysis to action in seconds, saying “keep a human in the loop” is no longer enough.
The boundary between recommendation and action is getting thinner
OpenAI’s Daybreak initiative is built around exactly this workflow: discover a vulnerability, validate whether it’s genuinely exploitable, and move toward remediation, all inside a single agentic loop. Palo Alto Networks is using frontier cyber models the same way, to find exposures, test exploitability, and help prioritize what gets fixed first.
Both are explicit that the action step stays under human control. That’s the current design choice, not a technical limitation. The industry has already built the discover-validate-remediate loop. What’s left is the governance layer that decides, in advance, what the system is authorized to do without asking.
What the data already shows
Research from the Cloud Security Alliance and Token Security, published in April 2026, found that 65 percent of organizations had already experienced a security incident caused by an AI agent operating on their network. Not projected. Already happened.
A separate survey found that when organizations were asked who is accountable for agent actions, most named someone. When asked how that accountability actually works in practice, the answer was usually informal or undefined.
Why “human in the loop” is too vague
The phrase hides four different jobs. Watching the system. Approving a recommendation. Authorizing an action. Overriding something that’s already started.
A security analyst may understand a technical recommendation perfectly and still not have authority to take a production system offline. A CISO may own cybersecurity risk and still not have unilateral authority to interrupt a critical business process. Operations may understand the consequence of an action without owning the control that triggers it.
Putting a human somewhere in the workflow doesn’t resolve any of those boundaries on its own.
Oversight means a human is watching. Authorization means a human with the right decision authority can approve, deny, stop, or escalate the action. Those are not the same thing.
Permission is not authority
An AI agent may have credentials that technically allow it to change a production environment. That’s a permission question. Whether the organization has actually authorized that agent to make that decision, under those conditions, is a governance question. Those are not the same thing, and most security programs have only answered the first one.
Five questions that expose the gap
What can the system observe on its own? What can it recommend? What can it act on inside a defined boundary? What requires explicit human authorization? And what is prohibited regardless of technical capability? If those five don’t have fast, documented answers, the system already has more authority than the organization does.
What this means for your organization
The question for leadership teams is direct: if your AI-enabled security system had to act in seconds, which decisions have you already authorized it to make without you?
If the answer is unclear, the governance model is lagging behind the technology, and that gap will matter more as autonomous cyber defense becomes more capable. Machine-speed security creates a human-speed governance problem. The question is no longer simply whether a human remains involved. It’s whether decision authority has been designed to operate at the speed of the technology.
If you want to find out before your AI system does
Governance Continuity Assessment
A 90-minute session that maps where your organization’s decision authority, escalation, and authorization boundaries actually stand, with a written finding within 48 hours.
Ann Marie van den Hurk, MSc., APR is the founder of Mind The Gap Advisory and originator of the CrisisOS5™ Framework. She advises CISOs, General Counsel, Chief Risk Officers, and boards on decision authority and executive crisis readiness for the AI era. Based in Newport, Rhode Island — serving organizations in Providence, Boston, Portsmouth, Portland, and Hartford, and across New England, nationally, and globally. mindthegapcyber.com