Shadow AI: What It Is, Real Risks, and How to Detect and Control It

What IBM’s 2025 Cost of a Data Breach study found about shadow AI (Ponemon Institute, 600 breached organizations, breaches March 2024–February 2025):

Shadow AI examples

Shadow AI examples include pasting source code or customer data into a public chatbot, AI note-takers joining meetings, AI browser extensions, and AI features switched on inside apps the company already approved.

• Public chatbots on personal accounts. In 2023 Samsung engineers uploaded internal source code to ChatGPT, according to a company memo reported by Bloomberg; Samsung then restricted staff use of generative AI tools (CSO Online).
• AI meeting note-takers that an employee connects to their calendar, which then record and transcribe calls with customers.
• Browser extensions that read the page you’re on to summarize or rewrite it, including internal tools and email.
• AI features inside approved SaaS. A new “AI assistant” button in a tool you already pay for can send data to a model provider your vendor review never covered.
• Unsanctioned AI agents that employees connect to company accounts with OAuth access to read or send data.

Shadow AI vs shadow IT

Shadow AI is a kind of shadow IT: shadow IT is any technology used without IT’s approval, and shadow AI is the subset that is AI. The difference is what happens to the data. An unapproved file-sharing app stores your data; an AI tool may also process it, keep it and, depending on its terms, use it to train a model.

Why shadow AI is a problem

Shadow AI is a problem because sensitive data leaves the company with no contract, no review and no log, and because the cost shows up in breaches. In IBM’s 2025 study, organizations with high shadow-AI use paid $670,000 more per breach on average than those with little or none.

• Data leakage: customer records, source code or deal terms end up with a provider the company never vetted.
• Compliance: privacy laws, health-data rules and customer contracts limit where data may go; an unapproved tool can break them without anyone knowing.
• Intellectual property: trade secrets can lose legal protection if a company doesn’t keep them secret.
• Bad decisions: output from an unvetted model goes into reports, code or customer replies without review.
• No governance: 63% of breached organizations in IBM’s study had no AI governance policy or were still developing one.

The use is widespread. Microsoft and LinkedIn’s 2024 Work Trend Index, a survey of 31,000 people in 31 countries, found 75% of knowledge workers used AI at work and 78% of AI users brought their own AI tools. In a Gartner survey of 302 cybersecurity leaders (March–May 2025), 69% of organizations suspected or had evidence that employees used prohibited public generative AI, and Gartner predicts more than 40% of enterprises will have a security or compliance incident linked to shadow AI by 2030.

How to detect shadow AI

You detect shadow AI by looking where AI use leaves traces: network and endpoint traffic, sign-ups and OAuth grants, the browser, and data leaving the company. Each method sees something the others miss. The products below are examples with published facts, not a ranking.

Start with the logs you already have. Firewall, DNS and identity-provider logs often show the biggest AI domains without buying anything new.

How to control shadow AI (without banning AI)

The way to control shadow AI is to make the approved route easier than the unapproved one, then watch and block what’s left. Samsung’s 2023 response was a ban; a ban with no approved alternative doesn’t stop anyone using a personal phone, where you see nothing.

1. Approve tools that do the job. Buy business versions of the AI tools people already use, with contract terms on data retention and training.
2. Write an AI acceptable-use policy. Say which tools are approved, what data may never go into any AI tool, and who approves new ones.
3. Keep an AI inventory. NIST’s AI Risk Management Framework asks for “mechanisms … to inventory AI systems” (GOVERN 1.6).
4. Add technical controls. Use DLP rules for sensitive data and discovery tools to block high-risk AI sites.
5. Train people. Show real examples of what not to paste, and how to request a new tool.

Gartner gives CIOs similar advice: enterprise-wide AI usage policies, regular audits for shadow AI, and GenAI risk checks in SaaS vendor reviews.

Shadow AI, ISO 42001 and NIST AI RMF

ISO/IEC 42001 and the NIST AI Risk Management Framework both turn shadow AI into a governance task: know what AI you use, set a policy, and assess the risk. The NIST AI RMF (January 26, 2023) is voluntary and free; its GOVERN 1.6 subcategory covers AI inventories.

ISO/IEC 42001 goes further because it’s certifiable. It requires an AI management system with policies, risk and impact assessments and 38 Annex A controls, which you can’t run without knowing where AI is used. Our ISO 42001 explainer covers cost and certification, and GRC software can hold the AI inventory and policy evidence. Once you know which AI tools are in use, AI red teaming tests how they behave.