Shadow AI discovery: how six vendors describe finding unapproved AI

Traffic inspection recognises AI services in the traffic a security cloud or proxy carries. SaaS awareness adds AI features that arrive inside applications the company already uses. Endpoint and agent inventory lists AI apps, agents, coding tools and MCP servers on devices and in systems, including those that never send traffic through the inspected path.

Cato Networks states it discovers shadow AI and coding agents across endpoints and auto-inventories every third-party AI tool, copilot and agent.

Source: Cato Networks: AI security · Read 7 Oct 2026

Source: Cato Networks: AI security for end users · Read 7 Oct 2026

Cisco states AI Access inspects web traffic to identify uses of shadow AI, including specific API calls made to AI models as part of AI-enhanced applications.

Source: Cisco Secure Access: AI Access · Read 7 Oct 2026

Fortinet states FortiAI lets you discover AI applications in use and understand their categories, use cases, risk levels and log details.

Source: Fortinet: AI security · Read 7 Oct 2026

Netskope describes continuous discovery of remote MCP servers, clients, tools, resources and prompt requests, and quotes its own research that 72% of generative AI use in the enterprise is shadow AI; general GenAI app discovery is not detailed on the pages we read.

Source: Netskope: Agentic Broker · Read 7 Oct 2026

Source: Netskope: securing generative AI · Read 7 Oct 2026

Palo Alto Networks states comprehensive visibility into shadow AI across GenAI applications, agents and marketplace plugins, by app and user, and gives catalogue counts of 4000+ GenAI applications and 80+ GenAI-specific attributes.

Source: Palo Alto Networks: AI Access Security · Read 7 Oct 2026

Zscaler states it detects and classifies thousands of AI apps, including AI embedded in popular SaaS applications, and its AI Asset Management page states it auto-discovers models, agents, MCP servers and shadow AI across 2,900+ apps.

Source: Zscaler: AI Access Security · Read 7 Oct 2026

Source: Zscaler: AI security · Read 7 Oct 2026

Two vendors publish catalogue counts, and they count on different bases. A large catalogue helps with recognising web AI services, but it does not help with an agent running on a laptop that talks to a model through a pinned connection, or an AI feature switched on inside a SaaS tenant. Ask which of the three discovery methods the product uses, and test with the AI tools your own people use, including the obscure ones.

A discovery report is useful for one meeting. After that, the work is sorting what you found into approved, tolerated and not allowed, and choosing a response for each. The quiz on this site starts with discovery as a worry, but it also asks what should happen next, because the response is what people will feel.

Which AI services, SaaS AI features, agents and MCP servers does discovery find, and through which method? How quickly are new AI tools added? Does discovery cover devices off the corporate network? Can discovery results feed policy directly, so a newly found tool starts in a watch or coach state automatically?

Track a few numbers each month: AI tools discovered, the share of them sorted into approved, tolerated or not allowed, requests for new tools and the time taken to answer them, and the number of watch, coach, redact and block events by category. A rising count of discovered tools is normal in the first months. A rising count of blocks on the same tool usually means people need an approved alternative, not a stricter rule. These numbers also show whether the discovery method is seeing what you expected, which is the real test of coverage.

Picking shadow AI as a worry adds points to the security cloud approach and triggers the check Finds shadow AI use. Five vendors state it on the pages we read and Netskope partly states it, because its stated discovery centres on MCP. The check is deliberately broad; the notes on each profile say what kind of discovery the vendor describes, so read them before you shortlist.