“The biggest gap is between daily AI use and what security can still see”: Vakaris Noreika, VP of Product Research and Development at NordLayer, on AI security

AI has quickly become part of everyday work, but many organizations still don’t have a clear view of how employees are using it. Chatbots may be covered by corporate licenses, while personal accounts, desktop AI applications, and coding agents can operate outside the systems that security teams traditionally monitor.

That creates a new challenge for AI security in the workplace: understanding not only which tools employees use, but what corporate data those tools can access through integrations and MCP servers, and what happens during an AI interaction. In an interview with Cybernews, Vakaris Noreika, VP of Product Research and Development at NordLayer, discusses the rise of shadow AI, the security risks of AI coding agents, and why visibility needs to come before stricter AI governance.

NordLayer AI Scout, a new AI security solution, was designed to close the gap between daily AI use and what security teams can see. Noreika also explains how companies can use it to see which AI tools employees use across browsers and desktops and what data they share, then turn that insight into practical rules rather than just banning AI tools.

Why doesn't a company's AI license list reflect real usage?

AI use inside companies is no longer limited to the tools IT teams have approved and licensed. Employees may use ChatGPT, Claude, or Copilot through personal accounts, while developers run coding agents that can access files and other tools directly from their machines. Others turn to AI services that the company never officially adopted.

That makes a company's AI license list a poor measure of actual usage.

“A license report shows the seats you pay for. The personal account, the desktop agent, and the text that went into the chat sit outside it,”

Noreika says. NordLayer AI Scout brings those different forms of AI activity into a single view, linking interactions to the people who use them.

What can an AI coding agent do that a chatbot can’t?

Coding agents and web chatbots can both access corporate data and tools, whether through integrations, MCP servers, or other connections. But coding agents can also operate directly in a developer’s environment. They can read repositories, change files, run commands, and interact with tools that the security team may not have reviewed. Sensitive information may be accessible to the agent without a developer explicitly pasting it into a prompt. As Noreika puts it,

“The sensitive file may never land in a chat box. The agent already has it, because it sits on the machine.”

Security teams, therefore, need visibility beyond the conversation. NordLayer AI Scout monitors the desktop sessions of supported coding agents, including conversations, tool calls, and the MCP servers and skills loaded during a session. It scans that text for sensitive data. On supported web-based AI chat platforms, it can also warn, redact, or block a prompt before it is sent.

What kind of sensitive data are employees exposing through AI tools?

The risk is not limited to one type of sensitive information. Customer records, financial data, source code, API keys, and internal documents can all pose problems when employees send them to AI tools without realizing the risk of exposure.

Rather than ranking one category as the most dangerous, Noreika suggests a simpler test:

“Would this person put that in an email to someone outside the company?”

If the answer is no, the same information should not automatically be considered safe to share with an AI service.

This is where AI DLP can help organizations detect sensitive information during the AI interaction itself. NordLayer AI Scout scans prompts for secrets, financial data, personal identifiers, and other patterns defined by the company, allowing teams to record, warn, redact, or block activity depending on the risk.

Why is shadow AI so difficult for security teams to detect?

Unapproved AI tools can be difficult for security teams to track because they often operate outside the systems that manage corporate software. An employee can open a chatbot with a personal account or install a coding assistant directly on a laptop without going through the company’s IT or identity systems.

“The corporate login system never sees it,”

he says. Though network logs can show that someone visited an AI service, and some network solutions can even attribute that traffic to specific employees, they might not provide a complete picture of AI tool usage.

That is why Noreika argues that shadow AI discovery needs to start on the device where the activity happens. NordLayer AI Scout’s browser extension and desktop sensor can identify the tools employees use and connect that activity to individual users, giving security teams a clearer picture of AI use beyond company-approved services.

Why should visibility come before restricting AI use?

Once organizations can see how employees are using AI, they can make better decisions about where controls are actually needed. Without that visibility, banning a tool or buying more licenses can amount to guesswork.

“AI observability lets a company learn the real pattern before it writes a rule,”

Noreika says. That means understanding which tools are being used, who is using them, and what is happening during those interactions, including whether sensitive data is involved.

For Noreika, the starting point is therefore observation rather than restriction. NordLayer AI Scout records AI activity so security teams can understand adoption and identify where stricter policies may be justified.

Why isn't browser-based monitoring enough for AI security?

AI security becomes harder when employees move between browser-based services and applications running directly on their computers. ChatGPT, Claude, Gemini, and Copilot may run in a browser or a desktop app, while coding agents such as Cursor and Claude Code can access files and other tools locally or in the cloud.

“A browser extension never sees that local session,”

Noreika says. As more capable AI tools operate on desktops, monitoring web activity alone can leave a significant portion of an organization's AI use out of view.

NordLayer AI Scout combines browser and desktop monitoring, giving security teams visibility into both types of activity rather than treating local AI applications as a separate environment.

What can AI usage data reveal that license counts can't?

A company may know how many AI seats it has purchased without knowing whether employees actually use them. That distinction can affect everything from software renewals to decisions about which tools should become part of the company's standard AI stack.

“A license shows what the company pays for. Usage shows what people actually do,”

Noreika says. NordLayer AI Scout can show which employees and teams use particular AI tools, how often they use them, and which models are involved, while also providing estimated usage-based spend.

That can reveal both underused paid tools and heavily used services that were never formally adopted, giving organizations a more realistic picture of workplace AI.

How can businesses monitor AI use without overreaching on privacy?

Greater visibility into AI use also raises questions about how much of an employee's interaction an organization should retain. Noreika says companies need to be able to choose the level of detail they collect rather than automatically storing every conversation.

NordLayer AI Scout offers different capture modes, ranging from keeping full prompts and detected sensitive values to retaining only the category of information detected, such as an API key or card number. Scanning occurs on the device, with excluded content removed before it is sent onward.

This gives organizations a way to investigate AI security risks while limiting how much potentially sensitive conversation data is retained.

How can organizations prove their AI governance is working?

As AI becomes part of everyday business processes, organizations need to be able to show not only that they have policies in place, but also how those policies are being applied.

“The question you will be asked is simple. How do you know how AI is used here, and what did you do about it?”

Noreika says. A policy document alone may not answer that. Organizations also need records showing who used which tools and when, what sensitive data was detected, and which controls were configured or changed.

NordLayer AI Scout keeps captured conversations alongside data-protection findings and records administrative changes with the user and time. Noreika also says the platform can send usage events to a SIEM and provides a quarterly evidence pack designed to help organizations document their AI use and controls.

How does AI agent autonomy change what governance needs to cover?

As AI systems become more capable, the security challenge is shifting from what employees type into an AI tool to what the tool can do with the access it has.

“The challenge moves from what the person typed to what the agent did in their name,”

Noreika says. Coding agents already illustrate this shift through MCP servers, skills, and tool calls that can extend an agent’s capabilities beyond the original conversation.

That poses a challenge for organizations that still govern AI primarily through approved tool lists. An agent may have the same permissions as an employee while taking actions that the employee did not explicitly describe. As these systems gain access to more business processes, security teams will need to understand which agent is acting, which tools it can access, and what data is moving through those interactions.

Where should businesses start with AI governance?

For organizations that still have little visibility into workplace AI, Noreika recommends starting with observation rather than restrictions.

“See the use first. Keep people working while you learn,”

he says. Monitoring activity across browsers and desktops can show which tools employees rely on and where sensitive information is appearing before a company decides which controls it actually needs.

Organizations can then introduce targeted rules for recurring risks, such as secrets or customer data, while allowing employees to continue using the AI tools they depend on. For him, practical AI governance comes down to visibility first, followed by evidence-based controls and regular reviews of how AI adoption evolves.

Closing the AI security visibility gap

AI adoption is moving faster than many organizations can keep up with. Employees are already using personal accounts, desktop AI applications, cloud agents, homegrown AI apps, and coding agents alongside the tools their companies officially provide, while security teams are often still relying on approved-app lists and license reports to understand what is happening.

For Noreika, the gap is ultimately about visibility.

“The biggest gap is between daily AI use and what security can still see,”

he says. The answer is not necessarily to restrict AI use, but to understand it well enough to decide where controls are actually needed.

That means knowing which tools employees use, what data moves through them, and what AI agents can access or do on a user's behalf. With that visibility in place, organizations can build policies around real behavior rather than assumptions, and adjust them as AI becomes a larger part of everyday work.

The challenge, then, is not simply keeping up with the latest AI tool. It is making AI use visible and governable before the technology moves further ahead of the rules designed to control it.

“Seeing it now is what makes a later rule possible.”