Stop Shadow AI Before It Stops You: A Practical AI Governance Playbook for Leaders

AI is spreading across your business faster than your policies. Teams are experimenting with local models on laptops, browser extensions, and free tools. That energy creates value, but unfortunately, it also creates a visibility and AI governance gap that puts data, operations, and compliance at risk. If leadership cannot see how and where AI is being used, you cannot manage it, measure it, or prove it is safe.

The Current AI Reality Is Local, Invisible and Unmanaged

AI use has escaped managed cloud services. Anyone can now run powerful models locally and with zero oversight. It is convenient, but it can lead to sensitive data being compromised. Not all external models or services respect your privacy or compliance policies. A single uncontrolled AI implementation can expose data outside of your governance perimeter.

At the same time, regulators are raising the bar. The EU AI Act, data protection authorities, and industry frameworks expect organizations to know what AI they use, document it, and control it. Executives do not need to be experts in every clause, but you do need proof that your program meets recognized standards and that controls actually work.

Why Are IT Leaders Losing Visibility?

Most companies have shadow AI. Teams adopt browser-based tools and install local apps or test models on spare servers. They rarely pass through the appropriate procurement channels or the CMDB. As a result, AI tools go unreported, undiscovered, and thus unmanaged.

Devices that can run AI, because of their CPU, GPU and memory, are especially important to identify. That is why step one is a unified view of your technology estate. Without a unified asset view, you will miss where AI is running, what data it can touch, and who is accountable.

You cannot govern what you cannot see. Start with an asset map that shows:

• AI-capable devices across endpoints and servers
• Installed local AI tools and artefacts
• Browser-based AI usage and extensions

Regain Visibility to Reclaim Control

This is where Lansweeper helps. Lansweeper aggregates and normalizes asset data across your environment so you can act on a single, accurate view. Two new Lansweeper reports make shadow AI visible in minutes:

• AI-Capable Assets Report identifies IT assets with the hardware specifications to run AI locally.
• AI-Active Assets Report identifies IT assets that are currently running AI locally.

Use the first to understand potential exposure and set guardrails. Use the second to focus controls and remediation where AI is already in use.

Extending Visibility into the Browser

Many AI risks now originate in the browser. Employees install extensions or use web apps that move sensitive text and files through external services. To address this, we are integrating traffic sensor capabilities into Lansweeper, following the Redjack acquisition.

Over the next few months, our team will reshape RedJack capabilities into new Lansweeper functionality that will give executives and security teams clearer views of browser-based AI application use. The goal is to surface which users, devices, and business units are interacting with AI tools in the browser. Based on that data you can coordinate or instigate informed policies to allow, guide or (where necessary) restrict usage.

Agentic Browsers: A Next-Wave Risk

Lately, a new category of AI-powered browser tools – often called agentic browsers – are emerging. These have the capability to not just summarize or assist, but to act autonomously on behalf of the user (e.g., navigate sites, fill forms, access connected services). Examples include Comet from Perplexity AI.

Recent security research shows that, unfortunately, these tools introduce novel attack surfaces:

• The browser may treat web content and user instructions without distinction, allowing attackers to embed hidden commands inside page content.
• In the so-called “CometJacking” attack, a single crafted URL can trigger the browser to access user memory or connected services (e.g., email, calendar), encode data (e.g., base64) and exfiltrate it, all without traditional phishing.
• Traditional web-security guards like same-origin policy and CORS become ineffective when the AI browser is executing commands across domains under the user’s logged-in context.

What Does This Mean for Your Team

Any AI-browser tool capable of autonomous actions must be treated like a privileged application.

• You need visibility.
  Gather information on which devices are running such browsers, whether they are connected to corporate accounts, what privileges they hold, and what governance/training controls exist.
• Explicitly address agentic browsers in your AI policy.
  Require registration, vetting, usage logs, limit automatic actions to human-approved mode, restrict connections to sensitive data or services unless fully audited.
• Thoroughly assess new AI-browser tools.
  Verify how they separate user intent from web content, how they handle connected services, how they log actions, and how they support enterprise control or blocking of risky features.

A Practical Operating Model for Better AI Governance

Policies are a great basis for AI governance, but your team needs to day-to-day concrete actions to properly implement them. Translate frameworks into an operating model your teams can run:

• Govern.
  Name an executive owner, define decision rights, and set minimum controls for data, security and auditability. Require registration for any AI system that processes sensitive data.
• Map.
  Catalog each AI system’s purpose, inputs, outputs, data sensitivity and stakeholders. Link systems to the assets they run on and to the business processes they support.
• Measure.
  Track reliability, security posture, privacy risks and model limits with simple metrics. Use a standard intake checklist for new use cases.
• Manage.
  Prioritize risks, fix issues, test changes and document outcomes. Start with assistive AI that proposes actions and requires human approval. Expand autonomy only after you have evidence of accuracy and safe behavior.

Quick Wins Executives Can Sponsor

Defining your policies and setting up your framework and processes can take some time. AI is already here. Here’s what you can do to take back control today:

• Shadow AI amnesty.
  Invite teams to register AI tools in exchange for support and clear guardrails. Seed your catalog with these submissions.
• Block and bless.
  Approve a small set of tools for common tasks and restrict clearly risky options. Publish simple playbooks so people have a safe path to productivity.
• Critical asset controls.
  Using Lansweeper’s AI-Capable and AI-Active reports, enforce patching, endpoint protection, and segment anything running local models until verified. Focus especially on data flows: where are you sending data, is it being used to train external models, who has access, what logging exists.

Where to Start

You do not need a 50-page policy to make progress. You need a clear line of sight into AI use, simple guardrails, and an operating rhythm that scales with demand. Turni AI governance from an abstract ambition into concrete steps

• Inventory. Run Lansweeper’s AI-Capable Assets and AI-Active Assets reports to produce your first view of potential and current local AI usage. Share results with asset owners.
• Policy. Issue a concise AI use policy and a one-page registration form. Make registration mandatory for any use that touches customer, employee, financial or other sensitive data.
• Controls. On AI-capable and AI-active devices, enforce baselines: patching, EDR, MFA, logging and network segmentation. Require allow-lists or code signing for local model runtimes. Also, explicitly audit agentic browser installations and their connection to privileged accounts or services.
• Operate. Stand up a small cross-functional AI review board. Use the Govern–Map–Measure–Manage loop for every new use case. Tie approvals to risk and business value.

Start with visibility, convert it into control, and grow from there. Lansweeper provides the asset intelligence to move quickly and confidently, while the Redjack integration is designed to bring your browser-based and agentic-browser usage into view and help coordinate policy decisions.