What is shadow AI? Why the browser tab matters
Employees can expose company data through unsanctioned AI tools. This guide explains the risks and the limited role of browser isolation.

Shadow AI is AI tool use at work without approval. A browser tab is all it takes. An employee pastes code or customer data into a chatbot. On an unmanaged device, the company may see nothing.
Policy, identity, and data controls address it. Browser isolation changes only where the page runs. It does not inspect what people type.
What counts as shadow AI
Shadow AI is unapproved AI use at work. It covers chatbots, AI browser extensions, and copilots. Approved, governed use is not shadow AI.
IBM's breach report calls it unregulated, unauthorized AI use. The term borrows from shadow IT. That older term covers unapproved software and cloud services. Shadow AI follows the same pattern. The difference is what happens to input.
Why employees turn to unapproved AI tools
Shadow AI happens for ordinary reasons. The tools are useful, fast, and easy to reach.
- AI tools live in the browser. No install and no IT ticket. An employee opens chatgpt.com and starts pasting.
- Free plans skip procurement. Claude and Gemini both list free plans. No one files an expense report.
- The payoff is immediate. Faster drafts and quicker answers reward each use.
- Governance lags adoption. Policies trail the tools people already use.
- Personal devices sit outside controls. Home laptops on home networks often skip corporate network tools.
Bring-your-own AI is common at work. A 2024 Microsoft and LinkedIn survey covered 31,000 people. They worked in 31 countries. It found 78% of AI users bring their own tools. That figure dates from May 2024.
Shadow AI also shows up in breach data. IBM's 2025 study covered 600 breached organizations. One in five reported a breach due to shadow AI. Only 37% had AI governance or detection policies. Ponemon Institute ran it, with IBM as sponsor.
The real risks of shadow AI
The risk is not that employees use AI. The risk is data going in without controls.
Data leakage through prompts
Pasted code or customer records go to the provider's servers. What happens next depends on the provider and the plan. Anthropic's August 2025 consumer terms are one example. Free, Pro, and Max users choose whether chats train models. Opting in extends retention to 5 years. Commercial plans fall under separate terms.
In May 2023, Samsung temporarily banned generative AI tools. The ban covered company computers. Bloomberg reported that staff had uploaded sensitive code to ChatGPT.
Even encrypted AI chats can leak signals. The Whisper Leak analysis explains how.
Sensitive information in AI prompts walks through common cases.
Intellectual property exposure
Trade secrets and unreleased plans can end up in prompts. Once sent, the company no longer controls them alone. The legal questions are still unsettled. The exposure itself is concrete.
Compliance obligations
Regulated data carries handling rules. HIPAA governs patient health data. PCI DSS governs payment card data. GDPR governs personal data of people in the EU. An unapproved AI vendor may not meet those terms. Check before regulated data goes in.
No audit trail
Personal accounts leave the company without records. Without other controls, the company cannot tell what left. An investigation or audit then has little to examine.
Unreviewed output
Output carries risk too. AI-drafted code or text can reach production unreviewed. Errors then ship with it.
Why browser-based AI use is hard to see
Shadow AI needs nothing more than a browser tab. An employee opens a chatbot and pastes in data. The data leaves over an encrypted HTTPS connection. Basic network tools see the destination, not the prompt.
Visibility depends on where a control sits. Microsoft Purview's endpoint DLP watches managed devices. Microsoft's cloud app discovery reads network and endpoint logs. A personal laptop on home Wi-Fi feeds neither.
The browser is where data enters the AI tool. Controls in the browser can see the paste itself. Purview, for example, can inspect prompts typed in Edge. The AI security guide library covers related browser risks.
What common controls see and miss
Each control sees a different slice of shadow AI. The table compares seven common ones.
| Control | What it covers | What it misses |
|---|---|---|
| Acceptable use policy | Which tools and data are allowed | Enforcement, which depends on people |
| Domain blocking | Known AI sites on company networks | New tools, personal devices, and home networks |
| Cloud app discovery | AI apps seen in network and endpoint logs | Apps outside its catalog and unmanaged devices |
| Endpoint DLP | Sensitive pastes and uploads on managed devices | Personal devices and unsupported browsers |
| Identity and SSO | Work sign-ins to approved AI apps | Personal accounts on unapproved tools |
| Extension allowlist | Which browser extensions can install | AI sites used in an ordinary tab |
| Browser isolation | Where the AI page runs | What a person types or uploads |
Layering covers more of the gaps. Policy is necessary but not sufficient. Without enforcement, it relies on people choosing to comply. Blocking AI domains has limits too. New tools keep appearing. A full block also removes legitimate uses.
How organizations respond
Microsoft's guide to shadow AI starts with discovery. Blocking comes later, once usage is understood. A practical sequence looks like this.
- Discover what people use. Filter app discovery for generative AI. Review each app's risk.
- Approve a short list. Give people a fast, sanctioned option. A slow approved path invites workarounds.
- Block unsanctioned apps. Tag them unsanctioned, then block them at the gateway.
- Restrict AI browser extensions. Use an allowlist, such as Chrome's extension policies.
- Limit sensitive pastes and uploads. Endpoint DLP can audit, warn, or block them.
- Measure and adjust. Track which tools people try. Review exceptions on a schedule.
What browser isolation does and does not do
Browser isolation changes where an AI page runs. It complements policy and identity controls. By itself, it provides no DLP or audit logs. It does not govern tools either.
- Choose the execution boundary. Shield opens a page in an isolated browser off your device.
- Keep routing distinct. Ghost loads a page through a private route. It does not isolate the page.
- Add controls separately. Identity, upload, and monitoring controls still do their own jobs.
- Know what disposal does. Closing Shield destroys the isolated browser session. It does not delete what a provider already received.
- Plan for prompt visibility elsewhere. Isolation alone does not show what people typed.
Some enterprise browsers bundle monitoring, DLP, and isolation. Those are separate capabilities. Check which ones a product actually includes. The browser isolation guide explains the basics.
The goal is a clear boundary, not an absolute guarantee. Browser isolation addresses page execution. The organization still owns authorization, retention, and monitoring. People still choose what they submit. How Legba browser isolation works covers the architecture.
Where Legba fits
Legba's Chrome extension has two modes. Shield is the one relevant here. Shield runs the page in a browser that is not on your computer, destroyed when you close the tab.
For shadow AI, that changes one thing. An AI page opened in Shield runs off the device. Shield is not DLP. It does not inspect what people submit. Pair it with DLP and identity controls.
Agent builders reach the same routing and isolation through a ready skill. See how to run OpenClaw safely for an agent example.
The extension costs $5 a month, or $50 a year, for a limited time. Agent plans are on the pricing page.
How we compared the controls
We compared controls by what each can observe or enforce. Product examples come from Microsoft and Google documentation. We did not test these products. Coverage varies by browser and device. Check your own vendor's documentation. Sources were checked on October 30, 2026.
FAQs.
Is using ChatGPT at work shadow AI?
How do companies detect shadow AI?
What is the difference between shadow AI and shadow IT?
Can browser isolation stop shadow AI data leaks?
Should companies ban AI tools entirely?
Continue the AI security deep dive
More on AI privacy, browser isolation, and agent safety.
What is the Whisper Leak attack on AI chatbots?
Whisper Leak infers AI chatbot topics from encrypted traffic. Microsoft researchers tested 28 models. See what routing and isolation change.
What is browser isolation? The complete 2026 guide
Browser isolation keeps web pages away from your operating system. Compare local, remote, and extension-controlled types and their limits.
How to run OpenClaw safely without giving an AI agent your laptop
OpenClaw permissions define its real blast radius. Review tools, files, credentials, browser data, network access, logs, and retention.
Free for 30 days. No card required.
Choose a browser boundary for AI tools
Legba is a Chrome extension with two modes. Ghost gives you a private browser route. Shield opens a page in an isolated browser off your device.