Shadow AI · 2026-07-24 · 4 min

Detecting shadow AI without a proxy or endpoint agent

In virtually every organization, employees use AI tools that IT never approved, from free-tier ChatGPT to Midjourney. The classic reflex is a network proxy or DLP solution. That works, but there is a lighter route that stands in days rather than months.

Why the classic approach is heavy

Network inspection via a proxy, CASB or DLP tool requires integration with the security stack, endpoint rollout and often TLS interception. For enterprises with a mature security team that is feasible; for SMB and midmarket it means a months-long project with license costs exceeding the AI spend itself. And it answers only half the question: it blocks traffic, but builds no governance.

The lighter route: the providers' own admin APIs

The major AI providers expose rich management APIs through their enterprise plans that most organizations never use:

  • OpenAI: Compliance API, Usage API and audit logs. Who uses what, at what cost, per team or custom GPT.
  • Anthropic: Usage & Cost API and Admin API. Tokens and costs per model and workspace, seats and Claude Code metrics.
  • Microsoft: Graph Usage API and Purview audit. Copilot activity per user and adoption metrics.
  • Google: Reports API and Gemini audit logs. Usage per organizational unit and feature.

Normalizing those sources into one data model creates visibility into actual AI usage: which department uses which system, what it costs, and which tools show up in the data that are not on your approved list.

What admin APIs see
+Business accounts per provider
+Usage and cost per team or department
+Seats, licenses and unused seats
+Audit logs and compliance status
+Which models and features are running
What they miss
–Free-tier usage on personal accounts
–Tools outside your provider list
–Browser extensions and desktop apps
–What exactly was in the prompt
Start with the left column: you already have that data. The right column is a deliberate next step, not a starting point.

What it won't show you

To be fair: free-tier usage outside business accounts stays invisible to admin APIs. Covering that too requires an optional browser extension or desktop agent, as a supplement and not a foundation. Start with the data that already exists.

From detection to governance

Detection is step one. The EU AI Act asks for what comes next: every detected system belongs in the AI register, gets a risk classification and, where needed, documentation. Merely blocking shadow AI also destroys the signal: employees use those tools because they have a need the approved stack doesn't cover.

PRODUCT
AxonGuard

AxonGuard connects the admin APIs of OpenAI, Anthropic, Microsoft and Google and automatically flags unregistered AI systems, no proxy or agent required.

View AxonGuard →