# Shadow AI false positives and blind spots

URL: https://shadowaidetectiontools.com/guides/shadow-ai-false-positives-and-blind-spots/
Updated: 2026-11-16

Detection tools get three things wrong: they flag approved tools under new domains, they miss apps outside their catalog, and they cannot see use on unmanaged devices or personal accounts. The vantage point decides what stays invisible: network tools miss local agents, endpoint tools miss SaaS grants.

## What do detection tools get wrong?

They get catalog matches wrong in both directions. A new domain for an approved tool is flagged as new, and a new AI tool outside the catalog is missed. Microsoft states that apps outside the Defender for Cloud Apps catalog cannot be discovered by default.

They also mislabel the account. A domain log cannot tell a personal from a corporate login, so a tool that counts users by domain can overstate corporate use. Products that read the browser or identity record can separate them, and Harmonic and Bay both describe personal-account detection.

Sources: [Cloud app discovery overview, Microsoft Defender for Cloud Apps](https://learn.microsoft.com/en-us/defender-cloud-apps/set-up-cloud-discovery); [Harmonic Security pricing](https://www.harmonic.security/pricing); [Bay: browser platform](https://www.bay.io/platform/browser?utm_source=shadowaidetectiontools.com&utm_medium=referral&utm_campaign=guide); [Harmonic Security source (harmonic-security-29)](https://www.harmonic.security/solutions/shadow-ai-detection).

## Which kinds of AI use stay invisible?

Five stay invisible to some tools. Use on an unmanaged phone or home laptop reaches no corporate sensor. Use inside an approved app's new AI feature leaves no new domain. Local models and CLI agents leave no network trace. Pasted text inside a permitted AI app is visible only to tools that inspect content. Offline use leaves nothing.

Each of those has a matching vantage point: a browser extension for personal-account web use, SaaS APIs for embedded features, an endpoint agent for local models and CLI agents, and content inspection for pasted text.

Sources: [Harmonic Security: endpoint AI security](https://www.harmonic.security/solutions/endpoint-ai-security); [Zscaler: public AI inventory, embedded AI](https://help.zscaler.com/ai-asset-mgmt/about-public-ai-inventory-embedded-ai); [Push Security: shadow AI, how to discover, govern and secure AI apps](https://pushsecurity.com/blog/shadow-ai-how-to-discover-govern-and-secure-ai-apps).

## Where does each vantage point go blind?

Network and proxy tools are blind to local software and to devices off the network. Browser-only tools are blind to desktop apps and CLI agents. Identity and email tools are blind to anything that never authenticates through the provider. Endpoint agents are blind to unmanaged devices.

Product files bear this out: Nudge Security says it does not touch the corporate network or endpoints, WitnessAI is network-only, and Push Security's vantage is the browser. Choose a second source that covers the first one's blind spot.

Sources: [Nudge Security product](https://www.nudgesecurity.com/product); [WitnessAI Observe](https://witness.ai/observe/); [Push Security: shadow AI, how to discover, govern and secure AI apps](https://pushsecurity.com/blog/shadow-ai-how-to-discover-govern-and-secure-ai-apps).

## How do you test a tool's coverage?

Plant known cases and see which are found: a test AI web app on a managed laptop, a personal-account login in a browser, an extension, a coding agent with an MCP server, and an OAuth grant on a test user. Record which tool found which and how long it took.

Run the test in a proof of value, with the same cases for every candidate. Vendors state times from minutes to 48 hours; your cases show the real figure.

Sources: [Reco: shadow AI discovery](https://www.reco.ai/use-cases/shadow-ai-discovery); [Harmonic Security: shadow AI detection](https://www.harmonic.security/solutions/shadow-ai-detection).

## How should the limits be reported?

Report limits next to the numbers: the sources used, the devices covered and the kinds not covered. A board count of "37 AI tools" without a coverage line suggests completeness the data cannot support.

State the share of managed devices with an agent, and the identity tenants connected.

## Which fixes reduce false alarms most?

Three fixes cut false alarms the most: maintain an approved list so known tools are matched, tag corporate IP ranges and test accounts, and connect an identity source so sign-ins resolve to people and tenants. The first two are policy work; the third needs an integration.

Check that the product separates the approved version of a tool from the consumer version. The same brand can appear as a licensed enterprise tenant and a personal login, and only a source that reads the account tells them apart.

Sources: [Bay: browser platform](https://www.bay.io/platform/browser?utm_source=shadowaidetectiontools.com&utm_medium=referral&utm_campaign=guide); [Akamai: AI usage control](https://www.akamai.com/solutions/ai-usage-control).

## Sources

- [Cloud app discovery overview, Microsoft Defender for Cloud Apps](https://learn.microsoft.com/en-us/defender-cloud-apps/set-up-cloud-discovery) (learn.microsoft.com), read 2026-11-16
- [Harmonic Security pricing](https://www.harmonic.security/pricing) (harmonic.security), read 2026-11-16
- [Bay: browser platform](https://www.bay.io/platform/browser?utm_source=shadowaidetectiontools.com&utm_medium=referral&utm_campaign=guide) (bay.io), read 2026-11-16
- [Harmonic Security: endpoint AI security](https://www.harmonic.security/solutions/endpoint-ai-security) (harmonic.security), read 2026-11-16
- [Zscaler: public AI inventory, embedded AI](https://help.zscaler.com/ai-asset-mgmt/about-public-ai-inventory-embedded-ai) (help.zscaler.com), read 2026-11-16
- [Push Security: shadow AI, how to discover, govern and secure AI apps](https://pushsecurity.com/blog/shadow-ai-how-to-discover-govern-and-secure-ai-apps) (pushsecurity.com), read 2026-11-16
- [Nudge Security product](https://www.nudgesecurity.com/product) (nudgesecurity.com), read 2026-11-16
- [WitnessAI Observe](https://witness.ai/observe/) (witness.ai), read 2026-11-16
- [Reco: shadow AI discovery](https://www.reco.ai/use-cases/shadow-ai-discovery) (reco.ai), read 2026-11-16
- [Harmonic Security: shadow AI detection](https://www.harmonic.security/solutions/shadow-ai-detection) (harmonic.security), read 2026-11-16
- [Akamai: AI usage control](https://www.akamai.com/solutions/ai-usage-control) (akamai.com), read 2026-11-16

## Related pages

- [Reading Proxy and DNS Logs for AI Apps](/guides/reading-proxy-and-dns-logs-for-ai-apps/)
- [Personal vs Corporate AI Accounts: Telling Them Apart](/guides/personal-versus-corporate-ai-accounts/)
- [traffic fingerprinting (glossary)](/glossary/traffic-fingerprinting/)
- [All guides](/guides/)
- [Shadow AI Discovery From Identity Provider Logs](/guides/shadow-ai-discovery-from-identity-provider-logs/)
- [Shadow AI Metrics for the Board](/guides/shadow-ai-metrics-for-the-board/)
- [The ranking](/ranking/)
- [Glossary](/glossary/)
