Data Loss Without an Attacker: Governing What Employees Put Into AI

On May 5, 2026, a U.S. regional bank discovered that its non-public customer information had been handled through an unauthorized AI model. Customer names, social security numbers, and dates of birth were among the information disclosed. The bank determined the event to be material because of the volume and sensitive nature of the information involved.[1] The resulting filing was the first under Form 8-K triggered by unauthorized AI use rather than an external attack.[2] An employee’s misuse of AI was placed in the same disclosure category as a cyberattack. The case is not isolated. IBM’s 2026 Cost of a Data Breach Report found that the share of security incidents involving an AI model increased from 13 percent to 21 percent this year.[3]

This raises a question many security programs have not yet asked: what does a DLP policy see when the data leaves as a prompt?

Why an Approved AI Tool Was Not Enough

Traditional data loss prevention tools were designed around files, endpoints, and network traffic. A prompt is different. It is text typed or pasted into a browser session, and the sensitive content within it rarely arrives as a labeled file.

The exposure is not limited to banking. In most organizations, what goes into a prompt is copied straight from everyday office documents: a sales rep summarizing a customer list in Excel, a legal team reviewing a contract in Word, or a manager condensing a strategy deck in PowerPoint. Once that content is pasted into a browser, it leaves behind the file and any protection that came with it.

Controlling the Path and the Prompt

Fasoo AI-R DLP gives employees a secure, approved path to generative AI, monitors the data they send and applies policy before sensitive information leaves the organization.

• Secure access through Remote Browser Isolation (RBI). Employees reach public generative AI services only through an isolated browser session set by the organization. Teams keep the productivity of public AI, while every session stays inside a monitored, policy-controlled path.
• Detailed blocking policies. Administrators can set specific blocking policies based on the types of data they do not want employees sending to generative AI.
• Pattern matching combined with AI detection. Proven pattern-matching methods work alongside AI technology to detect sensitive information such as PII, confidential business data, or proprietary information.
• User notification at the point of violation. A pop-up message blocks the action and tells the user which policy was violated, correcting behavior at the moment of the attempt.
• Comprehensive log for activity tracking. The log gives administrators a holistic view of all user activity, including the location, type, and time of each user’s access to AI tools.

Conclusion

This filing shows that prompt-based exposure can now be reported as a material cybersecurity incident, in the same category as an external attack.

Organizations that govern how employees reach AI and what they send to it, rather than only which AI is permitted, are positioned to adopt AI without waiting for the next security leak to occur.

See how AI-R DLP gives employees safe access to generative AI and prevents sensitive data from leaving through prompts. Book a demo.

Sources & References

[1] CB Financial Services, Inc., Form 8-K, Item 1.05 Material Cybersecurity Incidents, filed May 2026: sec.gov

[2] Wilson Sonsini, “Shadow AI Triggers First SEC Form 8-K for Unauthorized AI Use: What Financial Institutions and Public Companies Need to Know,” May 2026: wsgrdataadvisor.com

[3] IBM and Ponemon Institute, Cost of a Data Breach Report 2026, July 2026: ibm.com