Get a free audit
Apple credit Apple credits the researcher who runs our AI penetration testing.

Field Notes / Enterprise Security

Shadow AI in 2026: the statistics, the risk, and a policy you can copy

Half your staff use AI tools nobody approved. One in five organisations has already been breached through them, at $670,000 extra per breach. Here are the verified numbers, what shadow AI exposes, and an eight-clause policy you can adapt today.

Author
Red Team Partners
Read
11 MIN READ
Filed
13 March 2026
Updated
10 Nov 2026
An employee working late with a laptop and printed documents, using AI tools no one on the security team approved.

01 Shadow AI statistics 2026

You already run an AI estate. Most of it arrived without a purchase order or a security review, and it now touches email, source code, customer records and finance. Each figure below links to the publisher's own report, so you can quote it to your board.

49% Employees using AI tools their employer has not sanctioned BlackFog / Sapio, Jan 2026
1 in 5 Organisations breached through shadow AI IBM Cost of a Data Breach 2025
$670K Extra breach cost where shadow AI use is high IBM Cost of a Data Breach 2025
37% Organisations with a policy to manage AI or detect shadow AI IBM Cost of a Data Breach 2025
39.7% AI interactions that involve sensitive data Cyberhaven, Feb 2026
65% Shadow AI breaches that exposed customer personal data IBM Cost of a Data Breach 2025

Read the first and fourth numbers together. Half the workforce uses unapproved AI. Fewer than four in ten organisations have any policy to manage it, and only 34% audit for unsanctioned AI at all IBM 2025 . The rest find out about a tool when something goes wrong with it.

The stakes rose again this year. IBM's 2026 study puts the average breach at $4.99 million, up 12%, and finds one in four malicious breaches is now AI-enabled IBM 2026 . Attackers use AI to find the gaps faster. Shadow AI hands them more gaps to find.

02 Why employees use shadow AI

Your people use unapproved AI because it makes them faster and nobody gave them an approved option. The same BlackFog research led with a blunter finding: 60% of employees said they would take a security risk to meet a deadline BlackFog 2026 . Deadlines win against a policy nobody has read.

The UK's NCSC put it plainly in September: "You cannot manage what you do not know" ITPro 2026 . A blanket ban drives use onto personal phones and personal accounts, where you see even less. The firms that shrink shadow AI give staff a sanctioned tool that is as good as the one they found themselves, then set clear data rules around it.

The hardest cases come from the most productive staff. Non-developers now run coding assistants to scrape leads, push CRM updates against live databases and automate invoice runs. Each of those jobs grants the tool a real credential. Compromise that one person's project configuration and an attacker inherits everything they could reach.

03 What it exposes you to

Shadow AI lets data out through four doors. On engagements we check all four, because closing one pushes use to the next.

Pasted data. Staff paste customer records, contracts or code into a personal account. The data leaves your control with no data-processing agreement behind it. Samsung banned public chatbots in 2023 after engineers pasted semiconductor source code into one.

Connected tools. Agents, plug-ins and browser extensions connect through OAuth grants and API keys that skip security review. Each grant is a standing door into email, files or CRM. Prompt injection, the top risk in the OWASP Top 10 for LLM applications OWASP 2025 , lets a malicious document or web page instruct that agent to send your data elsewhere. AI agent security testing shows which of those grants an injected instruction can actually reach.

Unvetted suppliers. Every shadow tool is a supplier you never assessed: unknown hosting, unknown retention, unknown subprocessors. When that supplier is breached, your data is in the haul and you are not on the notification list.

Regulatory exposure. Personal data sent to an AI provider without a contract is a GDPR and UK GDPR failure the day it happens. The EU AI Act sets fines of up to €15 million or 3% of global turnover for breaching its obligations, and up to €35 million or 7% for prohibited practices EU 2024/1689 . Sector rules stack on top: FCA, HIPAA, FINMA. When shadow AI is involved, IBM found 65% of breaches exposed customer personal data, against a 53% average IBM 2025 .

04 A shadow AI policy you can copy

A policy that says "do not use AI" gets ignored by Friday. This one fits on a page, names what is allowed, and gives staff a fast route to a yes. Copy it, fill in the brackets, and have your data protection lead check it against your contracts.

AI acceptable-use policy · template

1. Approved tools

Staff may use these AI tools for work, through company accounts only: [list]. Anything else needs approval from [owner] first. Personal AI accounts are not to be used for company data.

2. Data you may never enter

Customer or patient personal data, passwords and API keys, source code from [repos], unreleased financials, legal and HR case material, and anything marked Confidential. This applies to every tool, approved or not.

3. Data you may enter into approved tools

Public information, your own drafts, and internal material classed Internal or below. When in doubt, ask [owner] before you paste.

4. Asking for a new tool

Request it through [form or channel]. We reply within [5] working days. A yes comes with a data-processing agreement and settings that block training on our data.

5. Agents, plug-ins and extensions

No AI tool may be connected to email, calendars, CRM, code repositories or databases without written approval from IT. Browser extensions that read page content need the same approval.

6. Checking the output

You own anything you send, publish or ship, even if AI wrote it. Check facts, figures, code and legal wording before use.

7. Reporting a mistake

If you pasted something you should not have, tell [owner] within 24 hours. Early reports are treated as a near miss, not a disciplinary matter.

8. Review

This policy is reviewed every [6] months and whenever we approve a new tool.

Clause 7 does the most work. Staff who fear discipline hide mistakes, and a hidden paste is a breach you learn about from the regulator. A 24-hour, no-blame report turns it into a near miss you can contain.

05 Finding and closing it

You close shadow AI in three stages, and the first takes about two weeks. Inventory every tool from network logs, SSO and OAuth grants, expense reports and browser-extension audits. Map what data flows into each one. Then interview each department head, because the conversations surface tools the logs never show. At the end you can name your AI estate, which puts you among the 34% of organisations that audit for it at all.

Next, publish the policy above and back it with controls: AI-aware data-loss prevention that spots sensitive data heading for an AI service, a central gateway that logs every call, and hardened settings for the tools you approve. Your people keep the speed. You keep the audit trail.

Last comes proof, and it is where most programmes lapse. AI red teaming of your approved AI systems shows whether prompt injection, system-prompt extraction or a poisoned retrieval pipeline breaks them, and gives you the report a regulator asks for. For a single model or chatbot, an LLM penetration test against the OWASP LLM Top 10 is the narrower option. Configurations change every month, so a one-off test ages fast. RTP Robin lets your team re-run the same attacks as tools change, so the picture stays current between formal assessments.

Shadow AI action log
  • Inventory every AI tool from traffic, OAuth grants, expense and browser audits, then confirm with department interviews
  • Map the data flow for each tool: what goes in, where it is processed, what comes out
  • Publish a one-page AI acceptable-use policy with approved tools, banned data and a fast request route
  • Deploy AI-aware DLP and route AI calls through a central, logged gateway
  • Red-team the approved systems, then re-test whenever a tool or configuration changes

Questions readers ask

What is shadow AI?

Shadow AI is any AI tool your staff use for work without approval from IT or security. Typical examples: a personal ChatGPT account used to summarise customer emails, a browser extension that reads every page, or a coding assistant wired into a live database. It is the successor to shadow IT, with one difference that matters. AI tools process the data you give them and send it to a model provider you never vetted.

How many employees use shadow AI?

About half. A January 2026 BlackFog survey of 2,000 UK and US employees found 49% use AI tools their employer has not sanctioned. Cyberhaven measured that 39.7% of AI interactions involve sensitive data such as customer records, source code or financials.

What does shadow AI cost in a breach?

IBM's Cost of a Data Breach 2025 found that one in five organisations had a breach due to shadow AI. Where shadow AI use was high, breaches cost an average of $670,000 more than where it was low or absent, and 65% of those breaches exposed customer personal data against a 53% average.

What should a shadow AI policy include?

Eight things: the approved tools and accounts, data that may never be entered, data that may be entered into approved tools, a fast route to request a new tool, rules for agents and plug-ins that connect to company systems, ownership of the output, a no-blame route to report a mistake, and a review date. A ban on all AI gets ignored. A short list of approved tools with clear data rules gets followed.

How do we find the shadow AI already in use?

Combine technical discovery with conversations. Network and DNS logs show traffic to AI services, SSO and OAuth grants show connected apps, browser-extension audits and expense reports show paid tools. Then ask each department head what their team actually uses. Interviews regularly surface tools the logs miss.

Does shadow AI create regulatory exposure?

Yes. Sending personal data to an AI provider without a data-processing agreement is a GDPR and UK GDPR problem. The EU AI Act sets fines of up to €15 million or 3% of global turnover for breaching its obligations, and up to €35 million or 7% for prohibited practices. "We did not know staff were using it" is not a defence.

References

Sources

  1. IBM. Cost of a Data Breach Report 2025: press release, "13% of organizations reported breaches of AI models or applications". IBM Newsroom, 30 July 2025. newsroom.ibm.com
  2. BlackFog. Shadow AI threat grows inside enterprises (research by Sapio Research, 2,000 UK and US employees, fieldwork November 2025). BlackFog, 27 January 2026. blackfog.com
  3. Cyberhaven. Sensitive enterprise data is flowing into AI tools at scale. Cyberhaven, 11 February 2026. cyberhaven.com
  4. IBM. Cost of a Data Breach Report 2026: "One in four malicious breaches are AI-enabled". IBM Newsroom, 29 July 2026. newsroom.ibm.com
  5. ITPro. "You cannot manage what you do not know": the NCSC is calling for a crackdown on shadow AI. ITPro, 8 September 2026. itpro.com
  6. OWASP. Top 10 for Large Language Model Applications, 2025 Edition. OWASP Foundation, 2025. owasp.org
  7. European Parliament and Council. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 99. Official Journal of the European Union, 2024. artificialintelligenceact.eu