What is shadow AI?
Generative AI usage inside an organization that security and IT never approved or provisioned — and why the usual shadow-IT signals don't catch it.
Shadow AI is generative AI usage inside an organization that security and IT never approved or provisioned — ChatGPT open in a personal tab, a free-tier tool one team bookmarked. Cyberhaven's 2024 research puts it at 70%+ of enterprise AI use, and shadow AI detection requires watching the browser itself, since classic shadow-IT signals (SSO logs, CASB catalogues) never see it.
Last updated:
The definition
Shadow AI is generative AI usage inside an organization that security and IT never approved or provisioned — ChatGPT open in a personal tab, a free-tier tool one team bookmarked, a self-hosted model nobody logged. It exists because these tools need no procurement, no install, and no admin rights to be useful. Every incentive points toward using them, and the friction that normally creates a paper trail — a purchase order, a security review, an IT ticket — simply isn't there.
Why it's different from classic shadow IT
The usual ways security discovers an unsanctioned tool don't work here.
A classic unsanctioned SaaS app is at least discoverable — it shows up in SSO logs, generates its own network traffic, or appears on an expense report. An AI tool opened in a tab of an already-approved browser, on an already-approved device, triggers none of those signals. A CASB catalogue is always a step behind, because new AI tools appear weekly and a catalogue can only list what's already in it.
What visibility typically turns up
Consistent categories, even though the exact volumes vary by organization.
- More tools than expected — usage spans well beyond the one or two tools on an approved list, including ones nobody in security had heard of.
- Usage concentrated in a minority of staff, usually engineering and customer-facing teams with the most repetitive text work.
- Source code is typically the most common sensitive category — a function pasted to ask why it's failing, often under a client NDA.
- Secrets appear less often than code, but they're the highest severity: there's no legitimate reason for a credential to be in a chatbot prompt.
- Almost none of it is malicious — the pattern is overwhelmingly competent people moving fast, not people trying to cause harm.
Cyberhaven's 2024 research puts unsanctioned shadow AI at 70%+ of enterprise AI use — the single largest reason a DLP built for known, catalogued apps can't see most of what's actually happening.
The full argument, including how to read this data without overreacting, is in what the first week of shadow AI telemetry usually shows.
How to get visibility
You can't govern usage you can't see, and the browser is the one place every web-based AI tool has to pass through.
A network proxy only sees traffic on the corporate network; a laptop at home walks straight past it. A CASB only sees the tools in its catalogue. Interception at the browser sees a tool the moment someone opens it, catalogued or not — the first week of that telemetry usually answers "what are we actually using" better than any policy document.
Find out what your team is
already sending to AI.
The first week of telemetry usually answers that better than any policy document. No credit card, no commitment — you reach the founding team directly.
Want to see it on your own team before talking price? Join the Design Partner Program — free for 30 days