Guide · Checked against vendor sites, September 2026
What is shadow AI, and how do you find it?
The verdict
Shadow AI is any AI tool, agent, extension or MCP server that employees use without the organization's approval or visibility. The practical response is discovery first, then policy: find what is in use, approve what is useful, and control the rest, because blanket blocking tends to push use onto personal devices.
Part of the University: Basics track
What counts as shadow AI in 2026?
It is no longer only chatbots in a browser tab. It includes coding agents installed from a terminal, desktop AI assistants, AI browser extensions, local MCP servers added to an agent's configuration, and plugins or skills pulled from marketplaces. Harmonic Security, for example, says it finds AI tools running locally that bypass corporate networks, and Prompt Security says it discovers shadow MCP servers and unsanctioned agent deployments.
Why doesn't network blocking solve it?
Many AI tools run locally or talk to their providers over ordinary HTTPS. A local MCP server may never touch the corporate network at all. That is why most tools in this guide discover AI use on the device, through a browser extension, an endpoint component, or existing EDR and MDM tooling.
How do tools in this guide discover shadow AI?
- On the device through existing EDR or MDM: Bay, Noma Security.
- Through endpoint management rollout of their own components: Harmonic Security (Intune, JAMF, Kandji, Group Policy).
- Through a browser extension, endpoint agent or gateway: Onyx Security, Lasso Security.
- Through an existing endpoint platform: Koi (Cortex XDR), Prompt Security (SentinelOne Singularity).
- Fleet-wide endpoint inventory including extensions and packages: Bloom Security.
What should happen after discovery?
Sort what you find into approved, approved with conditions, and blocked. Approved-with-conditions is where most value sits: a coding agent allowed for engineering, with shell commands that touch production credentials set to ask first. Tools that support ask or warn actions (Bay, Onyx Security, Harmonic Security) make that middle tier workable.
Frequently asked questions
Is shadow AI a security problem or an IT problem?
Both. It is a data and access risk for security and a sign of unmet demand for IT. Treat discovery results as a list of tools to approve as well as block.