Breaches tied to “shadow AI” doubled this year. Here’s how to protect your business

Chances are someone on your team has pasted a customer email, a spreadsheet, or a contract into a free AI chatbot this month. They weren’t being careless. AI helps them get through the day, and nobody told them which tools are safe to use.

Security researchers have a name for that: shadow AI. This year’s numbers show why it deserves a few hours of your attention.

What the research found

IBM’s Cost of a Data Breach Report, published July 29 and based on 602 organizations that had a breach between March 2025 and February 2026, found:

• Shadow AI incidents more than doubled. Security incidents involving shadow AI, where workers use AI their employer hasn’t approved, rose to 43% from 20% a year earlier.
• They cost more. Those incidents averaged $5.39 million, up from $4.63 million. Nearly half (49%) led to data being lost or exposed, and in about one in five the organization paid a regulatory fine.
• Most had no rules. 68% of breached organizations had no AI governance to manage AI or detect shadow AI.
• AI tools lacked basic locks. Among organizations with an AI-related breach, 92% lacked proper AI access controls, such as role-based access and multifactor authentication.

A small business’s numbers would be smaller than IBM’s averages, but the pattern applies to a ten-person office as much as to a large company: sensitive data goes into tools nobody chose, and nobody can see where it went.

The assistants themselves can be tricked

On August 18, Microsoft patched a flaw in Copilot Personal that the security firm Varonis named CoSnitch. A single click on a crafted link could make Copilot search a user’s connected Gmail, Drive, Calendar, and OneDrive and send the results to an attacker. Separately, a web page that Copilot was asked to summarize could plant instructions in its memory that survived password changes.

Varonis found no evidence that the flaw was used in real attacks. But the lesson applies to any AI assistant or agent connected to your accounts: it can be steered by what it reads. Varonis’s advice: “Treat Copilot as a privileged insider.”

Why banning AI doesn’t work

People use AI because it saves them time. Shut it off without an alternative, and the use moves to personal phones and accounts, where you can see even less.

Privacy worries are also what holds many businesses back. In the New York Fed’s regional business surveys, more than a third of firms not using AI cited data privacy, security, or confidentiality. In a Census Bureau survey, 20.7% of businesses not planning to use AI cited privacy and security concerns. The answer to both problems is the same: give people a safe way to use AI.

Five steps for a small business

1. Write a one-page AI policy. List the approved tools, what must never go into them (customer details, health or financial records, passwords), and when AI output needs a second look before it goes out.
2. Provide a good approved tool. A business plan from a major provider with proper data protections, or private AI on your own hardware, so the safe choice is also the easy one.
3. Limit what AI can reach. Connect assistants and agents only to the accounts and files they need. The fewer connections, the less a trick like CoSnitch can reach.
4. Require approval for anything sensitive. Sending money, deleting records, or emailing customers in bulk should wait for a person to say yes.
5. Train your team on the basics. Ten minutes on what not to paste, and on why an unexpected link that opens an AI assistant deserves the same suspicion as any other unexpected link.

One more thing: where AI runs doesn’t replace these steps. IBM found that AI-related breaches happened at about the same rate whether the AI was delivered as a cloud service or installed on a company’s own servers. Keeping AI in-house keeps your data in-house, but access controls matter wherever AI runs.

Where we can help

An AI policy, a short training session, and a plan for which tools to approve are part of our AI strategy and training work. It’s a small investment next to the cost of a breach, and it lets your team keep the time AI saves them.

Sources

• IBM: Cost of a Data Breach Report 2026 (PDF) and press release, July 29, 2026
• Varonis: CoSnitch: When Your AI Assistant Becomes Its Own Whistleblower, updated August 18, 2026
• Federal Reserve Bank of New York: Businesses Are Using AI to Transform Work, Not Cut Jobs, September 1, 2026
• U.S. Census Bureau: Business Trends and Outlook Survey, AI supplement (spreadsheet), 2026