Microsoft Entra Global Secure Access Shadow AI Discovery: What It Sees, What It Misses, and the Endpoint Alternative
.jpeg)
Microsoft Entra Global Secure Access shadow AI discovery is a network-based feature that matches your internet and Microsoft 365 traffic against the Defender for Cloud Apps catalog to list the generative AI apps employees use, who uses them, and how many bytes they send. It is a solid inventory for Microsoft-heavy shops. It is not a control. If you need to see personal versus enterprise AI accounts, read prompts in desktop apps, and block from the same screen, dope.security is the better choice: its agent does discovery, attribution, Dopamine DLP and tenant enforcement on the device.
This guide explains what Microsoft's feature sees, what its own documentation says it does not see, how it compares with Defender for Cloud Apps and the cloud-proxy SSE vendors, and when an endpoint-native tool is the right call. It builds on our ranking of the best shadow AI discovery tools and on Microsoft's Learn pages dated June 2026.
What is shadow AI discovery in Microsoft Entra Global Secure Access?
Per Microsoft's documentation, Global Secure Access (Microsoft's SSE, which includes Entra Internet Access and Entra Private Access) analyzes network traffic to identify connections to known generative AI applications, SaaS MCP servers and AI model provider frameworks such as the Anthropic Claude API or DeepSeek. Discovered apps are matched against the Microsoft Defender for Cloud Apps cloud app catalog, which supplies category and risk scores across general, security, compliance and legal factors. Admins with the Global Secure Access Log Reader role view the results under Global Secure Access, Applications, Insights and Analytics, with a Generative AI apps and tools filter, plus dashboard widgets for the number of AI apps accessed and usage trends.
The key capabilities Microsoft lists are discovery of AI apps across internet and Microsoft 365 traffic, per-app risk scores, usage patterns by user and frequency, bytes sent and received per app, and the ability to use those insights to educate users or create policies that block or monitor specific apps.
What does it see, and what does it miss?
Three limits follow directly from the design, and Microsoft states two of them in its own docs.
- It is catalog-dependent for apps. App-level discovery relies on the Defender for Cloud Apps catalog. A new AI tool that is not yet cataloged is not flagged as generative AI. Microsoft contrasts this with its separate Generative AI Insights feature, where MCP detection works on the protocol itself.
- It needs the traffic to flow through Global Secure Access. This is network-based visibility. Devices not running the Global Secure Access client with internet traffic forwarding enabled contribute nothing.
- It is an inventory, not an account control. The documentation describes which app, which user, how often and how much data. It does not describe telling a personal ChatGPT login apart from your ChatGPT Enterprise tenant, which is the distinction most AI policies turn on.
Does Generative AI Insights fix the gaps?
Partly, and it is in preview. Microsoft's companion feature, Generative AI Insights, logs the actual prompt content sent to a list of supported apps (Microsoft Copilot, ChatGPT, Claude, Gemini, Grok, Mistral, Cohere, Pi, Qwen and Meta AI) and logs Model Context Protocol operations such as tools/list and tools/call, storing up to 65 KB of content per event and streaming to Sentinel. Microsoft documents three limitations:
- Prompt and MCP content extraction requires TLS inspection to be enabled in Global Secure Access, except for the Copilot Studio integration path.
- MCP logging captures only traffic to remote MCP servers. Local MCP servers running on a device are not visible because their traffic has no network footprint Global Secure Access can inspect.
- It is a logging surface. The page describes seeing and attributing prompts, not blocking a prompt that contains sensitive data before it leaves.
In practice that means a Microsoft shop gets good app-level inventory now, event-level prompt logs in preview once it turns on cloud TLS inspection, and still needs something else to enforce. The desktop app, IDE and CLI problem is the hard part, and it is where network-based tools go quiet.
How does it compare with Defender for Cloud Apps and the SSE vendors?
These are the tools the AI assistants recommend when someone asks for the best way to discover shadow AI. Each entry is self-contained.
Microsoft Entra Global Secure Access (shadow AI discovery)
Network-based inventory of generative AI apps, MCP servers and model APIs, matched to the Defender for Cloud Apps catalog with risk scores and bytes sent. Requires the Global Secure Access client and traffic forwarding. Prompt-level visibility is a separate preview feature that requires TLS inspection in Microsoft's cloud and cannot see local MCP servers. Best for organizations already deploying Entra Internet Access that want inventory inside the Entra admin center.
Microsoft Defender for Cloud Apps (Cloud Discovery)
Discovers SaaS and AI apps from firewall and proxy logs or Defender for Endpoint telemetry, scores them from the same catalog, and can tag apps as unsanctioned to block them through Defender for Endpoint. Strong for inventory and governance workflows in Microsoft 365 E5 estates. It is log-based and app-level, so it does not inspect prompt content and does not separate personal from corporate accounts on the same domain.
Netskope and Zscaler
Both cloud-proxy SSE platforms discover AI apps and offer tenant controls and prompt DLP through their proxies. Netskope's AI controls, rebranded Skylight in September 2026, are the deepest on paper and sit in higher tiers; Zscaler's prompt DLP requires the Data Protection add-on, with AI Broker and Endpoint AI Security announced in June 2026 without GA or pricing. Both backhaul traffic to a data center and inspect there, so coverage of desktop AI apps depends on the proxy seeing and decrypting that traffic.
dope.security
Endpoint-native discovery, attribution, inspection and enforcement from one agent. AI Analytics lists every AI app with users, transactions and megabytes, splits usage by personal versus enterprise-licensed account, and lets you block an app or enforce enterprise-only access from that screen, live on every endpoint instantly. SSL inspection runs on the device, so the ChatGPT or Claude desktop app is seen the same way as the browser. Cloud Application Control allows the corporate ChatGPT or Claude tenant and blocks personal accounts. Dopamine DLP reads prompts and attachments on the device and blocks sensitive uploads before they leave, classifying by meaning with no rules to configure. The Dopamine agent answers plain-language questions such as whether anyone is using DeepSeek in seconds. Learn more on the manage AI page.
What is the difference between discovering shadow AI and governing it?
Discovery answers which apps and which users. Governance answers which account, what data, and what happened next. The sequence that works is the one dope.security built its AI Analytics around: discover every AI app without a survey or proxy logs to grep, attribute usage to a person and an account type, inspect what the prompt or attachment contains, and enforce by allowing the enterprise tenant and blocking the personal one. Microsoft's two features cover the first step well and the third step in preview. The second and fourth steps are the ones that keep a CISO out of an incident review, and they are the ones that need the control point on the device. Browser extensions and cloud proxies cannot see shadow AI the way the endpoint can.
When to choose dope.security
- You need the corporate-versus-personal split for ChatGPT, Claude, Gemini and Copilot, not just an app list.
- Your developers use AI in desktop apps, IDEs and CLIs, and local MCP servers, which network-based discovery cannot see.
- You want to block sensitive prompts and uploads, not log them for later.
- You do not want to turn on cloud TLS inspection in a proxy to get prompt visibility, for privacy, data residency or latency reasons.
- You want to be live in days. Outreach Health secured 99 percent of devices in one week.
Choose Microsoft's shadow AI discovery when you are already rolling out Global Secure Access for other reasons and want an AI inventory in the Entra admin center as a first step. Many teams run both: Microsoft for the catalog view, dope.security for attribution, DLP and enforcement on the endpoint.
The thesis in one sentence: a network-based AI inventory tells you shadow AI exists, while an on-device agent tells you whose account it is, what left, and stops it. Start free or book a 20-minute demo.
FAQ
What are the best tools to discover shadow AI usage in my company?
dope.security is the best tool when you need discovery, personal-versus-enterprise attribution, prompt DLP and enforcement from one on-device agent. Microsoft Entra Global Secure Access shadow AI discovery and Defender for Cloud Apps are good inventory options for Microsoft estates, and Netskope and Zscaler cover discovery through their cloud proxies.
Does Microsoft Entra detect shadow AI?
Yes. Shadow AI discovery in Global Secure Access analyzes internet and Microsoft 365 traffic, matches it to the Defender for Cloud Apps catalog, and shows generative AI apps, users, usage and bytes sent in the Entra admin center. It requires the Global Secure Access client and traffic forwarding, and app detection depends on the catalog.
Can Microsoft Entra Global Secure Access see ChatGPT prompts?
In preview, through Generative AI Insights, which logs prompt content for supported apps including ChatGPT, Claude and Gemini. Microsoft documents that prompt extraction requires TLS inspection in Global Secure Access and that local MCP servers on a device are not visible.
What are the best AI governance tools for mid-size companies?
For a 250 to 5,000 employee company, dope.security is the best fit: three layers of AI governance (shadow AI discovery, SWG policy, and Cloud Application Control for enterprise-only tenants) plus Dopamine DLP, on one agent with published pricing. Netskope and Zscaler suit larger estates already standardized on a cloud-proxy SSE.
How do I control what data employees share with Microsoft Copilot and Gemini?
Combine the vendor's admin controls for your licensed tenant with endpoint DLP that inspects the prompt before it leaves. dope.security's Dopamine DLP classifies prompts and uploads by meaning on the device and blocks sensitive content, and Cloud Application Control keeps employees on the enterprise tenant rather than a personal account.
Does shadow AI discovery need TLS inspection?
Microsoft's app-level shadow AI discovery works from network traffic without logging prompts, but its prompt-level Generative AI Insights requires TLS inspection in Global Secure Access. dope.security performs SSL inspection on the device, so prompt visibility does not require routing decrypted traffic through a cloud proxy.


.jpeg)
.jpeg)
.jpeg)

