Shadow AI is the use of AI tools, extensions and agents by employees without the knowledge or approval of IT or security. It is a form of shadow IT, but with a different risk: what you enter leaves your organisation as data. It also almost always starts with an identity you don’t manage, such as a personal account, a free subscription or a personal password.

Why does shadow AI happen?

Employees use AI because it saves them time. The tools are often free, run in the browser and need no installation. Summarising a text, checking code or writing up a meeting takes a few seconds.

Other causes:

There is no approved alternative, or employees don’t know about it.
There is no AI policy, so nobody knows what is and isn’t allowed.
Vendors add AI features to software that is already in use, often enabled by default.

Why shadow AI is an identity problem

Every interaction with an AI tool starts with an identity: a human or non-human entity with credentials, tokens or API keys. Many organisations manage access to SaaS through an identity provider (IdP) with SSO. Consumer AI doesn’t ask for a corporate login. An employee signs up with a personal address or a free account, and that usage stays invisible to your IdP.

As a result, you can’t answer:

What data employees share with those tools.
Which integrations, plug-ins and connectors are active.
Whether access to a specific tool should exist at all.

A corporate email address doesn’t solve this. An account on a free plan is often registered with a work address but sits outside your control. Reused passwords make it worse: if a login leaks elsewhere, the AI account and everything in it is exposed too. When an employee leaves, these accounts stay active if the offboarding process doesn’t know about them.

Without visibility into identities, you can’t monitor, restrict or revoke usage.

Examples of shadow AI

AI chatbots with a personal account or free subscription
AI note-takers and meeting bots that record and summarise meetings
AI browser extensions, which often ask for broad permissions on the pages an employee visits
AI assistants that get access to mail, calendars or files via OAuth
AI features in SaaS applications that nobody deliberately enabled
Self-built apps and automations using AI coding assistants or low-code platforms, connected through personal accounts or API keys
AI agents that connect to external tools via MCP servers

Risks of shadow AI

Data leaks. Customer data, contracts, source code or internal strategy reach an external party through prompts and uploads. You can’t take that back afterwards.
No control over accounts. A personal account has no SSO, MFA policy or logging. You can’t see who shares what, and you can’t revoke access.
Risky extensions. An AI extension runs inside the user’s trusted browser session. With high permissions, it can read prompts, web pages and clipboard content, and read or write in business applications.
Agents inherit permissions. An AI agent works with the access of the person who deploys it. If that identity isn’t managed, you can’t manage the agent either. Some MCP servers can also run commands, read and write files and connect to the internet.
A new form of shadow SaaS. Employees can use AI to build an app themselves that connects to customer data or internal documents. Such an experiment can quickly become a business process without IT knowing about it.
Privacy and compliance. Without a data processing agreement and insight into where data is processed, you can’t demonstrate GDPR compliance. Without an inventory, you also don’t know which AI systems fall under the EU AI Act or NIS2.
Unreliable output. AI can make mistakes. Results that are copied unchecked into quotes, code or customer communication create operational and legal risks.

How do you deal with shadow AI?

Banning it rarely works. A ban without an approved alternative pushes usage to channels where you see even less. Work in this order: visibility first, then policy, then enforcement.

Make visible which AI tools are used, including with personal and freemium accounts. Combine DNS, proxy and CASB logs, browser data, OAuth integrations and conversations with departments.
Set an AI policy. Define which tools are allowed and which data must never go into a public AI tool. Keep it short and specific.
Offer an approved alternative. A business AI environment with SSO and clear agreements on data processing pulls a lot of usage out of the shadows.
Assess tools by risk. Look at the data that goes in, the access the tool gets and the vendor’s terms. Set a clear policy per tool:
- Allow: the tool is approved.
- Warn: the tool is permitted, but the employee gets a notice and a pointer to the approved alternative.
- Block: access is not allowed, with an explanation at the point of access.
Control what leaves your organisation. Detect and block sensitive data in prompts and uploads.
Include AI accounts in offboarding. This covers accounts employees created themselves.
Train specifically. Explain which data doesn’t belong in an AI tool and why.

Which solutions help?

No single solution covers shadow AI on its own. These components complement each other:

IAM and SSO: make sure approved AI tools run through corporate identities, with MFA and central access management.
SSE and SASE: visibility into AI websites and apps, with policy per user group, including outside the office network. A secure web gateway is part of this.
CASB: discovers AI apps in cloud traffic and applies policy per app, for example access only through corporate accounts.
Data Loss Prevention (DLP): detects sensitive data going to AI services, in prompts and uploads, and can block it.
DNS security: shows which AI services are being requested and can block categories or specific services.
SSPM: shows which AI assistants and apps have OAuth permissions in your SaaS environment.
Browser security: gives visibility into and control over browser extensions and web apps, including AI extensions with broad permissions.

Which combination fits depends on what you already have in place and which data you most want to protect.

Questions to start with

Do your security tools see AI logins made with personal credentials?
Have approved SaaS products recently gained AI features you haven’t assessed yet?
Do you know which browser extensions employees use for AI?
Does your offboarding process cover AI accounts that employees created themselves?
Do you know which employees have deployed AI agents or automations, and with which permissions?

About shadow AI and Nomios

Nomios helps you see which AI tools are used in your organisation and what risks they create. We advise on policy and governance, and help you choose and set up solutions for identity, access and data security.

Ecosystem