Glossary. Shadow AI.
What is shadow AI, and what should you check first?
Shadow AI is AI used at work without approval. What decides the risk is the data each tool can reach, so the register starts there for approved tools as well.
By Dipal Panchal, co-founder and CEO · Updated November 8, 2026
The tool matters less than its reach
Shadow AI is the use of AI tools, assistants or agents at work without the organization’s approval or knowledge, with company data pasted, uploaded or connected to them. Unlike shadow IT in general, the risk sits in the data a tool can reach more than in the tool itself. The first check on any AI tool, approved or not, is what it can read.
The usual forms are a public chatbot used from a browser, a browser extension granted access to email, and an AI feature a vendor switched on inside an approved application.
An AI-tool register, six rows
Tools in one company, reviewed on November 7, 2026. The Role or group column records approval and who uses each tool.
| Identity | Role or group | Object | Privilege | Last used | Decision |
|---|---|---|---|---|---|
| Enterprise AI assistant | Approved; licensed to all staff | Every Microsoft 365 file the user can open | Reads with each user’s own permissions | Daily | Keep; fix overshared HR and finance libraries before wider use |
| Public chatbot, free tier | Not approved; finance team, UK | Whatever staff paste or upload: spreadsheets, PDFs | Uploads by users | November 4, 2026 | Block on the corporate network; point users to the approved assistant |
| Meeting-notes browser extension | Not approved; sales team, India | Calendars and email bodies | OAuth consent granted by each user | November 6, 2026 | Revoke the consents; require admin approval for new apps |
| Coding assistant | Approved; engineering | Source repositories, including config files | App installed on the code host | Daily | Keep; remove the keys found in config files first |
| support-bot, an in-house agent | Approved pilot; support | Every CRM table, through a service account | Database SELECT on the whole crm schema | November 7, 2026 | Narrow the account to the three columns the bot uses |
| AI feature in the HR system | Not assessed; switched on by the vendor | Employee records already in the HR system | Built into an approved application | Not measured | Excluded from the block list; assess under the vendor contract first |
Where unapproved tools show up
Unapproved tools leave records in places security already holds. OAuth app consents in your identity provider show extensions like row 3. Proxy or DNS logs show traffic to AI services like row 2, and expense reports show the subscriptions nobody filed with IT.
Rows 1, 5 and 6 show why the register covers approved tools too. Each has approval, and each reaches more data than its job needs until someone narrows the access.
The rules that ask for a register
None of these uses the term shadow AI. Each asks for something an AI-tool register provides.
| Rule | Provision | What it asks |
|---|---|---|
| NIST AI Risk Management Framework 1.0 | GOVERN 1.6 | Mechanisms to inventory AI systems, resourced according to organizational risk priorities |
| EU AI Act | Article 4, as replaced by Regulation (EU) 2026/1744, Art. 1(5), in force from July 27, 2026 | Providers and deployers take measures to support AI literacy among staff and others using AI systems on their behalf |
| GDPR and UK GDPR | Article 28 | A processor handling personal data for you works under a contract with set terms |
Check the data before the model
Dipal Panchal, Sylox’s co-founder and CEO, made the point on the FutureGPT panel in June 2026: “The model is rarely the risk. The risk is the data feeding it and the access around it.” Row 5 is that case: an approved bot with a read grant on every CRM table.
On an advisory call in 2026 he called the upload route the common one. A company licenses an AI assistant for every employee, and staff upload spreadsheets and PDFs that hold company data the company does not know about. So for every tool in the register, record what data it can reach, and narrow that before you decide whether to approve it.
What a register can’t see
A register lists the tools you found. A personal phone used with a personal account, outside the corporate network and identity provider, leaves none of the records above.
It also can’t recall data already uploaded. Once a spreadsheet is pasted into a public tool, the vendor’s terms decide what happens to it, which is why row 2 is blocked outright.
IRIS maps who can reach sensitive data in supported sources, and an agent’s service account is one more identity in that map. Capabilities vary by edition.
What an agent’s database or Microsoft 365 account can reach: IRIS produces part of this, for supported sources that expose permission data.
Sensitive data in the files an assistant can open: IRIS produces part of this, in the sources you connect.
Which public AI tools staff use (your web proxy or CASB logs show the traffic): Not produced by IRIS.
Blocking uploads or revoking app consents: Not produced by IRIS.
Explore AI governancePull the latest app consents
Export the OAuth app consents from your identity provider for the past month and mark every AI tool among them approved, not approved or not assessed. The first register is done when each tool has a row with the data it can reach and a decision.
Before an approved assistant goes wide, review oversharing in the libraries it will read, then work through the checklist for AI assistant oversharing.
Official sources
- NIST AI Risk Management Framework (AI RMF 1.0), GOVERN 1.6, NIST. Opened November 8, 2026.
- Digital Omnibus on AI, Regulation (EU) 2026/1744: Article 1(5) and amended Article 113, European Parliament and Council. Opened November 8, 2026.
From sharing to AI governance
What is oversharing?
The step before approving an assistant: files that reach more people, and so more prompts, than the work needs.
What is shadow data?
Untracked copies of sensitive data, which any AI tool with access can then read.
What is AI governance?
The next step: the policies, roles and records that decide which AI is used, and on what data.
What security teams ask about shadow AI
It can be. If the vendor acts as your processor, Article 28 requires a contract; if it uses the data for its own purposes, you need a lawful basis to disclose it. A vendor outside the EU or UK also raises the transfer rules in Chapter V.
No. Blocking without an approved alternative can move use to personal devices. Block the tools that take uploads, and offer one approved assistant with its access narrowed first.
Yes, when a feature inside it was switched on without review, like row 6. Approval covers what was assessed, and a vendor’s later additions need their own review.
Protect your brand. Start with your data.
Bring the service accounts your AI agents use. We’ll show what sensitive data they can reach, in supported sources.