Do you know where your AI agents are?
Do you know what they can access?
Do you know what they're authorized to do?
These questions served as the foundation for a survey sponsored by Okta, which released its findings in a report, Global CISO Insights 2026: Identity security in the age of AI, this past week.
Fewer than half of the survey respondents said yes to these questions: 47%, 46%, and 45%, respectively.
That's a serious, but not surprising, admission from the people who are supposed to be reducing this kind of risk. And it points at a harder question underneath the ones Okta asked: if identity platforms can't fully account for agents at the point of login, what happens to visibility once a session is already underway?
How big is the shadow AI governance gap, really?
These numbers are about confidence, but the uncertainty isn't abstract. In Okta's research, 68% of CISOs said they see at least some unsanctioned use of AI or agentic AI in their organization — and when asked what's standing in the way of closing that gap, the answer isn't budget or headcount. It's visibility: a lack of visibility into AI identities is the top-cited barrier to securing AI, at 48% globally and 74% in the U.S. Okta doesn't say how CISOs arrived at that 68% figure, which is itself worth sitting with — a number describing a visibility gap, produced without a clear visibility mechanism behind it.
Neon Cyber's research on how employees actually use AI explains why that gap persists. It isn't a matter of a few holdouts: 70.9% of knowledge workers use AI tools daily, but 54.6% of workers use unsanctioned AI tools, which makes shadow AI a permanent feature of how work gets done, not a phase to wait out.
CISOs aren't underestimating shadow AI. They're accurately describing a gap they don't yet have the tooling to close.
Why can't identity governance see unsanctioned AI use?
Because it was never built to.
SSO, MFA, and IAM answer one question extremely well: is this the person they claim to be, at the moment they're asking for access? That's a real and necessary security control, and it's the reason identity has become the backbone of enterprise security over the last decade.
Identity governance works exactly as intended under two conditions: the app is known to IT, and the employee is using their corporate identity to access it. When both are true, SSO and IAM can confirm who accessed what, and — depending on the app — that activity gets logged somewhere.
The gap opens the moment either condition breaks, and shadow AI breaks both routinely.
An unsanctioned tool was never onboarded to your IAM program, so there's no SSO login, no directory entry, nothing for governance to attach to.
And a sanctioned tool goes just as dark the moment an employee uses a personal account instead of a corporate one to reach it — same tool, same browser tab, now completely outside every identity control the enterprise has in place.
Neither case is a logging failure identity vendors overlooked. It's that identity governance can only govern what it's been told to look for, and unsanctioned tools and personal credentials are, by definition, not on that list.
In Quantifying Shadow AI Risk in the Browser, Neon Cyber highlighted how routinely both conditions break in practice: 49.3% of knowledge workers said they'd simply reach for a different unapproved tool if the one they wanted wasn't sanctioned. Blocking a tool doesn't close the gap, it just moves it — to some place your identity security program cannot, by definition, govern.
That workaround behavior converges on a single, specific blind spot: 42.3% currently run a sanctioned AI tool and an unsanctioned one in parallel, oftentimes in the same browser. Identity governance has visibility only into the sanctioned AI tool.
Identity governs known apps and corporate credentials. It has no way to see either the moment an employee steps outside them.
And with the proliferation of new AI apps and tools coming onto the market, this risk is only going to grow.
How shadow AI risk is expanding with AI agents
This is a critical insight from Okta's research: "AI agents have created a governance crisis at the heart of enterprise security". And "you can’t secure what you can’t see, and you can’t govern what you haven't given an identity to."
As employees start to use agentic AI for consumer tasks like finding and booking restaurant reservations or personal travel, they are going to realize how easy it would be to use AI agents for their work. Furthermore, as browsers enable and push agentic AI in the browser (already available with Google Chrome and Perplexity Comet), it's easy to see how agentic AI is the next logical step in AI use. From the network and identity security perspective, this all still looks like the employee. The login was real. The credentials were real. But the entity now acting inside the session might not be a human at all.
As agentic AI tools become easier to use and more embedded in everyday application workflows, the number of sessions where something other than the employee is taking the actions will keep climbing. If your enterprise has enabled specific AI tools across your IAM stack, you might be able to see what that agent does. But if you struggle to have visibility into shadow AI and shadow AI agents, that challenge will only grow.
From: Okta Global CISO InsightsThis is where in-browser telemetry is critical. Instead of relying solely on network traffic or identity governance, Neon Cyber sits alongside users in the browser. With AI Agent Visibility, currently available to select design partners, Neon can identify when a browser session is being driven by an AI agent rather than a human. The underlying signal is behavioral: AI agents act at a speed and cadence no human replicates, the same way CAPTCHA works but in reverse — instead of proving a human is present, the system looks for the absence of human-paced behavior as the tell that something else is now driving the session.
This is the kind of information that most CISOs can't currently answer about their own environment: not just who logged in, but what's actually acting once they're inside.
Identity governance needs shadow AI visibility to work
Security leaders who've invested in identity have done the right thing. But Okta's research highlights how identity security platforms are still catching up in an AI-driven world. Extending visibility from who authenticated into what's acting inside the session is simply an extension of this identity security conversation.
That's true even where identity governance for AI agents works exactly as intended. Governing an agent as a “first-class identity” — its own lifecycle, its own scoped permissions — still depends on knowing that agent exists in the first place. It has no mechanism for surfacing the unsanctioned AI tool an employee found on their own, or the AI agent operating inside a pre-approved application and a session that's been accurately authenticated. This isn't a one-time gap to close and move past — it's an ongoing risk as long as employees can discover and adopt new AI tools faster than security teams can review, assess, and onboard them to their identity governance platform.
This is where browser observability is critical. Neon Cyber monitors every click on a web page, data pushed into every prompt, and events across every application — sanctioned or not. Know when accounts are created, how they're authenticated, and where your identity perimeter is breaking down so you can strengthen it.
See Neon Cyber in action to learn how browser-native visibility strengthens identity governance and helps enterprises discover and manage shadow AI.
FAQs
What percentage of CISOs can identify all the AI agents in their environment?
Only 47% of CISOs surveyed by Okta in its Global CISO Insights 2026 report said they're confident they can identify all the AI agents operating in their environment. Confidence drops below 30% when limited to CISOs who "strongly agree" rather than "somewhat agree."
How does shadow AI use go undetected by identity security platforms?
Identity security platforms work when two conditions hold: the app is known to IT, and the employee is using their corporate identity to access it. Unsanctioned AI tools were never onboarded to identity in the first place, so there's no login event for governance to attach to — and even a sanctioned tool becomes invisible the moment an employee reaches it with a personal account instead of a corporate one.
What's the security risk difference between an employee using an AI tool and an AI agent acting on their behalf?
The risk isn't just that the actor is different — it's the speed and scale at which exposure can happen. An employee typing into an AI tool makes one decision at a time, at human pace, with at least some awareness of what's sensitive. An AI agent acting inside that same authenticated session can take dozens of actions in the time it takes a person to write one sentence, with no pause to weigh whether a document, field, or upload should leave the organization. Identity infrastructure may see a login, but it has no way to flag that the actor, and the risk calculus, has changed.
Does identity governance solve shadow AI on its own?
Not by itself. Treating AI agents as first-class identities — Okta's own recommendation — still depends on knowing an agent or tool exists before it can be governed. Neon Cyber closes that gap by monitoring authentication events across every browser-based app, sanctioned or not, including shadow tools that bypass SSO and never reach identity provider logs — so identity governance has something to govern in the first place.