Shadow AI Detection Tools

Embedded AI features switched on in approved SaaS

An approved SaaS app can add an AI feature, an assistant or a new sub-processor without a new domain or install, so network and endpoint tools see nothing change. Find these features through SaaS admin consoles, OAuth and permission changes, and vendor change notices, or with a tool that reads SaaS APIs.

Facts read on .

How do AI features appear inside approved apps?

They appear as a setting, a license add-on or an automatic update. The app, its domain and its sign-in stay the same, so nothing in a proxy log, DNS log or sign-in record changes. The data the app holds is now processed by a model, sometimes one run by a third party.

Nudge Security says it shows which SaaS vendors use third-party AI services such as OpenAI and Anthropic, which it frames as AI in SaaS supply chains. Grip Security says it flags approved SaaS that has AI features.

Sources: Nudge Security pricing; Grip Security: reduce shadow AI risks.

How do you find them?

Check three places: the SaaS admin console for AI settings and add-ons, the change notices that vendors send before enabling a feature, and OAuth or API permission changes tied to the app. For the largest suites, review the AI section of the admin center each quarter and after each major release.

Detection products read these records at scale. Zscaler documents a separate public AI inventory for embedded AI, Netskope says its AI Command Center covers embedded SaaS AI, and Harmonic Explore covers AI features embedded in SaaS tools. Reco and Grip Security list named embedded AI across their connected SaaS apps.

Sources: Zscaler: public AI inventory, embedded AI; Netskope AI Command Center documentation; Harmonic Security Explore; Reco: shadow AI discovery; Grip Security: discover and control shadow AI; Netskope source (netskope-06); Reco source (reco-04).

Which vantage points can see an embedded feature?

SaaS APIs and the vendor's own admin and audit records can see an embedded feature, because the setting lives in the app. Identity logs see it only when the feature needs a new consent. Endpoint and network views see it only when the feature calls a new host.

The Vantage Point Matrix on this site shows which products read SaaS APIs and logs. Microsoft documents SaaS posture sources in Defender for Cloud Apps, and Palo Alto Networks documents SaaS posture as one source for AI Access Security.

Sources: Manage app governance, Microsoft Defender for Cloud Apps; Palo Alto Networks: discover risks posed by GenAI apps.

What should the review decide?

For each embedded feature decide whether it is enabled for everyone, a pilot group or no one, which data it can reach, and where inputs go. Ask the vendor whether inputs train a model and how long prompts are retained; Nudge Security says it summarizes each AI vendor's model-training policy, opt-out options and retention.

Record the decision in the inventory against the app, not as a separate tool. An embedded feature is a property of an approved app and changes with it.

Sources: Nudge Security: AI discovery solution.

What controls help when a feature cannot be turned off?

When a feature cannot be turned off, limit the data it can reach. Use app permissions and sensitivity labels, restrict which groups get the add-on, and watch the paste and upload paths into the app. Microsoft documents Endpoint DLP detection of pasting content to restricted service domains in supported browsers, and Purview DSPM for AI detects sensitive information shared with AI apps.

Keep a dated record of the decision so an auditor can see when the feature was reviewed.

Sources: Endpoint data loss prevention, Microsoft Purview; Data Security Posture Management for AI, Microsoft Purview.

How do you keep the review current?

Subscribe an owner to each large vendor's release notes and admin-center message feed, and review again when a vendor announces an AI feature. Add AI settings to the quarterly access review so a feature enabled by default is caught within a cycle.

Nudge Security publishes a dated changelog of its own AI agent discovery coverage, which listed 23 agentic platforms on 2 November 2026; use dated vendor change logs like that to see which platforms a tool can read before relying on it.

Sources: Nudge Security changelog.

Sources