AI has handed security teams a genuine force multiplier—faster analysis, pattern recognition at scale, automated detection. It has also delivered the same capabilities to the people trying to break in. And it has introduced a third problem that tends to get less attention: Employees across every organisation are now routinely connecting to external AI systems, sharing data those systems retain and process, and most security programs can’t see it happening.
That convergence has raised the cost of visibility gaps considerably. But before addressing what’s new, it’s worth examining why the foundational problem—knowing what’s on your estate, what it’s doing, and whether that's normal—remains unsolved for so many organisations.
A fireside chat I hosted recently illustrated this. I asked a senior security practitioner what visibility meant to his organisation, and his answer came in two parts: what it meant when they started, and what it means now. Those two things were almost unrecognisable as the same concept.
From inventory to intelligence
The starting point for most organisations was asset inventory: what devices are on the network, who’s using them, and which software they’re running. That remains foundational—and yet is still unsolved in far too many organisations. And even for those making headway, the challenge continuously evolves to become ever more difficult.
My fireside interviewee described how his team now uses real-time endpoint telemetry to monitor devices during software update cycles—not just to confirm that an update has been completed, but to compare behavioural data in real time, establish what “normal” looks like at that moment, and use that intelligence to predict how similar rollouts will behave across other regions and configurations. That is not taking an inventory. It’s operational intelligence derived from visibility at a level of fidelity that simply wasn’t considered necessary a few years ago. The endpoint—where a user interacts with your systems, where data is created and accessed, where AI tools are run—is the most critical unit of this problem.
The estate that outgrew the map
Part of why foundational visibility remains elusive is structural. Digital estates in large organisations don’t grow neatly. They grow organically, shaped by budget cycles, project needs, and crises that force years of change into months.
I spent time working at the Ministry of Defence, which, in the early 2000s, ran what was by some measures the largest single IT network in the world. Even then, we were running network scans only every six weeks—not by preference, but because more frequent scans would have overloaded the infrastructure. A six-week-old picture of your estate in an environment where threats can materialise in hours is not a visibility capability. It’s a historical document. And that was before accounting for the corporal at a remote base who plugged a games console into a network port, or the smart TVs in meeting rooms that were still logged into internal Wi-Fi weeks after the last occupant had left.
<blockquote>A six-week-old picture of your estate in an environment where threats can materialise in hours is not a visibility capability. It’s a historical document.</blockquote>
COVID exacerbated this across sectors. Organisations that had never supported remote working went fully distributed almost overnight. Devices proliferated outside established visibility perimeters, and in many organisations the estate has never been fully remapped since.
And all of these challenges existed before AI ever entered the equation.
The AI visibility problem is three-fold
The first AI challenge is on the threat side. AI is enabling attackers to discover vulnerabilities that don’t appear in any existing database — found not by matching known patterns but by reasoning about code and systems directly. The exploitation window between a flaw being identified and being weaponised was already uncomfortably short. AI is narrowing it further, and doing so at a pace that outstrips most patch and response cycles.
The second issue is behavioural. Employees are connecting to external large language models — some sanctioned, most not — and in the process may be feeding confidential information into systems outside any governance or data protection framework. This is not a future risk. It is happening now, at scale, in organisations with mature security programs, and in most cases it is happening without meaningful visibility.
The third is internal. AI capabilities are now embedded in security and visibility tools themselves. That introduces a series of questions that organisations are only just beginning to ask: What model is being used, how accurate is it, and what does it do with the information it processes? Applying rigorous visibility standards to your estate but not to the tools governing that estate is a blind spot worth examining.
The patch gap and the third path
The average time from vulnerability identification to patch availability is still north of 90 days. Active exploitation often begins within days. For a growing category of zero-day vulnerabilities, the attack arrives before any patch exists.
That gap has traditionally left organisations with two options: accept the risk and continue operating, or remove the affected system and accept the operational disruption. Neither is satisfying. For critical systems, neither may be viable.
There is a third path, but it’s dependent on the quality of visibility. If you understand in real time how a specific vulnerability is being exploited—what the attack looks like at the endpoint—you can build targeted detections around that behaviour pattern before a patch arrives. You can then continue to operate, as you know what the exploit looks like in motion, and you can automate reactions at the point of attempt rather than after the fact.
Organisations doing this effectively are drawing on current threat intelligence to build custom detections at the endpoint level, constraining the blast radius and maintaining operational continuity during the patch window. That is a real capability — but it requires visibility with real-time fidelity, not weekly snapshots or scan cycles measured in weeks.
That capability also depends on something that often goes unexamined: confidence that your visibility tools are actually doing what they report. Many organisations assume this is covered, because EDR has been deployed, coverage reads at or near 100%; the dashboard is green. But EDR agents sometimes get disabled during troubleshooting and are not re-enabled. Or they fall out of date on machines that haven't been actively managed. In a recent proof-of-concept at a large organisation, we found that devices were operating without their standard EDR agent active, yet the EDR platform was reporting full deployment.
The gap between what a tool reports and what is actually true on the estate is its own category of visibility failure, and arguably the most dangerous kind: the blind spot you don't know you have. Continuous, independent discovery—verifying not just that assets exist but that the protections on them are present, current, and functioning—is not a supplement to a visibility regime. It is a core part of one.
Operating inside the gap
The 90-day patch gap exists for every organisation and AI has made operating inside it more consequential and more complex. What differs between organisations that contain incidents and those that become case studies is whether their visibility is real-time, complete, and trustworthy enough to act on when it matters.
The organisations managing this well have stopped treating visibility as a problem they solved and started treating it as a discipline that has to keep pace with an estate that never stops changing and a threat environment that is now actively using the same tools they are. The question you need to ask yourself is: Do you have a solution that is fit for the threat environment you're actually in?