What is shadow AI?
Shadow AI is any AI service, model, feature or agent used inside an organisation without having been approved through its normal governance process. It includes consumer chat services used for work, AI features switched on inside SaaS products, model API keys in developer code, and agents connected to company systems.
It is rarely malicious. People adopt AI because it helps them work, and the approval route is often slower than the tool. The problem is that each use sends data to a third party, or lets a model act, before anyone has decided whether that is acceptable.
Why it matters
The exposure can be measured. In the IBM Cost of a Data Breach Report 2025, 13% of organisations reported a breach of AI models or applications, and 97% of those lacked proper AI access controls. Where shadow AI use was high, it added USD 670,000 to the average cost of a breach.
Shadow AI is also a governance gap. A system nobody has recorded cannot be assessed, has no owner, and leaves nothing for an auditor to read.
Where shadow AI hides
| Source | Example | Why it is missed |
|---|---|---|
| Consumer AI services | A chat assistant used in a browser with a personal account | It is an ordinary HTTPS request to a popular domain |
| AI inside SaaS | A summarisation feature switched on in a CRM or a document tool | The vendor was approved before the feature existed |
| Developer API keys | A model API called from a script or an internal application | It looks like any other API dependency |
| Local tools | A coding assistant or a desktop AI client on a laptop | It runs on the endpoint, not in a central system |
| Agents and MCP servers | An agent connected to a ticketing system through an MCP server | The tools it can call are rarely written down |
How to detect shadow AI
No single method sees everything. The practical approach combines the sources you already have with one control point for the traffic you can route.
| Method | What it sees | What it misses |
|---|---|---|
| Proxy and DNS logs | Which AI domains are contacted, from which hosts, how often | Content, and anything off the corporate network |
| An AI gateway | Every routed request in full, with the decision made about it | Traffic that is not routed through it |
| Endpoint inventory | AI tools installed or running on managed devices | Personal devices and browser use |
| Declarations | Purpose, owner and data, from the people who use the system | Whatever people do not mention |
- Start with logs you already hold. Export a few weeks of proxy or DNS logs and match them against a list of AI providers. This gives a first map without touching any prompt.
- Agree a sanctioned list. Write down which providers and services are approved, so that everything else is, by definition, a finding.
- Route what you can. Point applications, SDKs and agents at one gateway, and use a PAC file or a proxy for traffic that nobody pointed there on purpose.
- Ask the people. A short declaration form catches purpose and owner, which no log contains.
From findings to decisions
Detection is only useful if every finding leads somewhere. Treat each one as a candidate record in an AI inventory, and give it a path.
- Confirm it. A person confirms that the finding is a real AI system in use, and records an owner and a purpose.
- Assess it. Ask what data can reach it, who reviews its outputs and what it can do. Classify the risk from those facts.
- Decide. Approve it with conditions and an end date, replace it with a sanctioned alternative, or block it.
- Enforce and record. Apply the decision where traffic passes, and keep the record of who decided, when and on what basis.
Blocking without a route to approval tends to push use out of sight. A short, visible approval path, with a sanctioned option people can use the same day, usually reduces shadow AI more than a longer block list.
A shadow AI checklist
- A written list of sanctioned AI providers and services, with an owner
- A first scan of proxy or DNS logs against known AI domains
- A declaration form for teams to record AI use, owner and purpose
- An AI inventory in which every finding becomes a record, observed or declared
- A defined route from finding to approval, replacement or block
- Controls on what data may be sent, applied where traffic is routed
- A review cycle, because new AI features keep appearing in approved tools
How MERIDVAR handles shadow AI
MERIDVAR is software you install and operate, so discovery runs on your own infrastructure and nothing is sent to us.
- Gateway comparison. Traffic is compared with your list of sanctioned providers, in off, observe or enforce mode.
- Log upload. Proxy and DNS logs can be uploaded to see services that never touched the gateway.
- Endpoint agents. Endpoint agents, with installers for Linux, macOS and Windows, report the local AI tools they find. The product is in beta.
- Inventory candidates. An observation never becomes a declaration on its own. A person confirms it, and it then follows the same assessment and approval path as any other AI system.
Controls apply to traffic routed through the gateway. Uploaded logs, endpoint agents and declarations extend visibility to the rest. See AI security for the admission controls and the product overview for the full path.
Source for the figures: IBM, Cost of a Data Breach Report 2025.