To detect shadow AI, start with records your organization already holds: DNS and web proxy logs, app consents in your identity provider, SaaS admin settings, device and browser inventories, expense records, and cloud and code accounts. Network records can show that a device connected to an AI service. Over an encrypted connection they cannot show what was typed, and they miss a personal phone on a mobile network. A no-blame staff survey helps cover what the records cannot.
Each record in the detection section of our shadow AI guide gets its own entry below: what it shows, what it misses and where to start. New to the term? Begin with what is shadow AI.
Start with what you hold
Records you already hold are the first place to look.
Know what the network misses
A log can show a connection to a service, not the prompt.
Agree the rules first
Legal, HR and privacy settle who may look before anyone does.
Finding is not blocking
Sort findings by the data involved, then govern.
Network records are where shadow AI detection starts, not where it ends. They can show connections to AI services, from which devices, and how often. Their limits:
The same sources, in the same order, as the guide's table of where to look for shadow AI.
DNS logs (DNS is the internet's address book) can show that a device looked up an AI service's name, and when: the lookup, not that anything was sent. A browser can look up a name without anyone opening the site, one more reason a lookup is only a lead. A web proxy or firewall sees the web traffic that is routed through it on its way out. Proxy and firewall logs can show which device, or which user where the proxy identifies users, connected to the service, how often and, where byte counts are logged, how much data moved.
Misses: encrypted content, devices off the network, and lookups that skip your resolver.
First step: search a fixed window of these logs for the names on your list. Count DNS lookups and proxy or firewall connections separately, and rank devices, and users where the proxy identifies them, by each. A high count is a lead to follow up, not proof of use. If these logs are not collected yet, turn collection on and run the search once a window has built up. The voluntary CIS Controls v8.1 list "Collect DNS Query Audit Logs" and "Collect URL Request Audit Logs" as Safeguards 8.6 and 8.7, each "where appropriate and supported."
When someone lets an outside app read their work mail, files or calendar, that is an OAuth consent, and identity providers such as Microsoft Entra ID and Google Workspace keep a record of the grant. By default, Microsoft Entra ID lets users consent to an app for permissions that don't require an administrator, such as letting an app read their own mailbox. Out of the box, Google Workspace lets users sign in with Google to any third-party app.
In Entra, the Permissions page of an application added through user or admin consent has Admin consent and User consent tabs, and the audit log records a "Consent to application" activity. In the Google Admin console, Security > Access and data control > API controls > Manage App Access lists the apps that have accessed Google data, with the number of users and the Google services each requests. A new app can take 24 to 48 hours to appear there. Each of these views needs an administrator role, which the linked pages name.
Misses: what an app did after the consent, and tools used with a personal account.
First step: download Google's list of accessed apps, review the enterprise applications in Entra, and start with AI tools that can read mail, files or calendars. Entra keeps audit logs for seven days on its Free edition and 30 days on P1 or P2, so pull them before they age out.
Platforms you already license can gain AI features through vendor updates, and the platform's own admin console is the place to look for them. Where it has AI settings, they may show which features are enabled and for whom; what is recorded varies by platform. Google's data retention table for Workspace, for one, lists "Gemini for Workspace log events", kept for six months.
Misses: who used a feature and what they typed, unless the platform keeps activity reports.
First step: list the platforms that hold your most sensitive data and note whether AI features are on, for whom, and whether your contract covers them.
Microsoft Intune's Discovered apps report lists the apps detected on Intune-enrolled devices (on personally owned devices, only the apps the organization manages). In the Google Admin console, the Apps & extensions usage report lists extensions on enrolled Chrome browsers and managed profiles once browser reporting is turned on. Both need administrator access to their console.
Look for AI browser extensions, AI apps and coding tools that install as apps, and apps that run AI models on the device itself, which may send nothing for network logs to catch. A coding assistant added as a plug-in inside a code editor may appear in neither report; the editor's own extension list, or the survey, is where to find it. An inventory shows what is installed, not whether it is used. We wrote about a featured Chrome extension caught harvesting AI chats.
Misses: devices that are not enrolled, browsers and profiles outside management, apps the organization does not manage on personally owned devices, web-only tools that install nothing, and plug-ins inside code editors.
First step: export the app and extension lists and search them for the names on your list.
Paid AI subscriptions on company cards or expense claims can show the vendor and who paid or claimed, which is not always who uses the tool. A subscription bought through an app store may show only the store's name.
Misses: free tools, and anything paid for personally.
First step: give finance the vendor names on your list and ask for matching charges in the same window as the logs.
Developers can reach AI models directly, through an AI service switched on in a cloud account or an API key (the credential a program uses to reach a service) in code. They can also connect AI tools to the organization's code repositories. GitHub's secret scanning has patterns for OpenAI, Anthropic and Google Gemini API keys. It runs free on public repositories; for an organization's private repositories it needs GitHub Secret Protection on GitHub Team or GitHub Enterprise Cloud. An alert shows that a key was exposed somewhere GitHub scans, not that it was used or that any data was sent.
Misses: keys kept where the scan does not look, and cloud accounts the organization does not know exist.
First step: review the AI services enabled and billed in each cloud account, check which apps and integrations are connected to your code platform, and run secret scanning where you have it. Revoke any exposed key at the provider; deleting it from the code is not enough.
Logs cannot see a personal phone on a mobile network or show what kind of information went into a tool. People can. Ask them in a short survey, under a no-blame promise agreed in advance with HR and legal:
Their answers are the first draft of your AI inventory.
The ground rules come before anyone opens a log.
Finding a tool is not, on its own, a reason to block it. Sort each finding by the kind of data that goes into it, by what the tool can reach (mail, files, code) and by whether it runs on business terms the organization has reviewed.
The terms matter. Whether an AI tool uses what you type to train its models depends on the product and its settings. OpenAI and Anthropic each say that, by default, they do not use inputs or outputs from their business products to train their models. Microsoft says prompts and responses in Copilot, as used by organizations, are not used to train foundation models. Google says Workspace customer data is not used to train models without the customer's permission. OpenAI, Google and Anthropic each say their consumer apps can use conversations to improve their models, depending on the user's settings. So record which account each tool runs under: a personal plan can be a consumer plan even when it is paid (OpenAI's Plus and Pro and Anthropic's Pro and Max are examples).
Patient, customer or financial data in a personal account comes first; public information in a reviewed business tool may need only a row in the inventory. Then govern, with the guide's AI governance basics and its case for why banning AI tools alone falls short.
Dedicated tools can do part of the searching, depending on the product and how it is set up:
A tool that reads what people type is reading employees' words, so the ground rules from step one apply to it in full.
Outside help fits when the records sit with teams that have no time to join them up, or when findings need to stand up in front of a board or an examiner. InfoSight's AI Governance & Exposure Assessment lists Shadow AI Discovery among its areas. The service page gives its focus as "What's actually running" and lists "SaaS/AI app inventory, browser extension analysis, DNS/proxy logs, CASB & Defender data, OAuth app audit."
New tools appear and vendors add AI features to products you already use, so a single sweep goes stale. Aim for a rhythm that fits your risk: for example, consent and extension reports monthly, the inventory each quarter, the survey once a year.
The voluntary NIST AI Risk Management Framework lists, under GOVERN 1.6, that "mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities." The NIST Cybersecurity Framework 2.0 lists, among its outcomes, that "inventories of software, services, and systems managed by the organization are maintained" (ID.AM-02). A tool no one approved is not on that inventory until someone finds it.
No. Network logs can show that a device connected to an AI service, when and how often. Over an encrypted connection they do not show what was typed, they miss personal phones on mobile networks, and they may not separate AI inside approved software from ordinary use. Add app consents, admin consoles, device inventories and a staff survey.
Not from ordinary network logs. Over an encrypted connection, they can show a connection to the service, not the prompt. For a tool the organization does not manage, seeing what was typed takes decryption at a gateway or a control in the browser or on the device. Either one means reading employees' words, so agree purpose, notice and who may look with legal, HR and privacy first.
Use the browser management you already run. In the Google Admin console, for example, the Apps & extensions usage report lists extensions on enrolled Chrome browsers and managed profiles once browser reporting is on. Search it for AI tools and check what each may read. Browsers and profiles outside management do not appear; a staff survey can help find them.
Treat it as a routine, not a one-time project. A reasonable aim is to re-run the app consent and browser extension reports monthly, review the AI inventory each quarter, and repeat the staff survey yearly. Check again when a major platform you use adds AI features, and adjust the rhythm to your risk and to how long your logs are kept.
Agree the ground rules with legal, HR and privacy: the purpose, the notice to staff, and who may look at records showing one person's activity. Then list the AI services to look for and search the records you already hold, starting with DNS and proxy logs and your identity provider's app consents.
Detection is the first step. The full guide to shadow AI covers what comes after it.
No audit language. No commitment.
Prefer the phone? Call 305.828.1003
Schedule a consultation from one of our expert advisors.
No spam—one security expert follow‑up, that’s it.
Stay informed of the latest cyber trends.