An employee pastes part of a contract into a chatbot to shorten the text. An accountant uploads a client spreadsheet to build a report faster. A developer shows the AI a piece of code with a database password in it. All of this is done to get work done faster, but the data ends up outside the company.
In its list of the top risks for large language model applications, OWASP singles out sensitive information disclosure: personal data, financial details and confidential business data. Below: how to avoid leaks through AI and where DLP fits in.
What the risk is
- Data leaves the company. What happens to it next depends on the service’s terms and the account settings — and the company has no control at all over employees’ personal accounts.
- Personal data. Passing clients’ or employees’ personal data to third-party services must comply with the law — see “GDPR for Ukrainian companies: when it applies”.
- Trade secrets and NDAs. Client contracts may explicitly prohibit passing their data to third parties.
- Secrets in code and documents: passwords, access keys, internal addresses.
- Excessive access rights. Corporate AI assistants work with company data. Microsoft, for example, states that Microsoft Copilot only surfaces data to which the user has at least view permissions. So if permissions are granted too broadly, the assistant can easily find and retell what the person was never meant to see.
What DLP is
DLP (Data Loss Prevention) prevents data leaks. In the definition NIST cites, it is a system’s ability to identify, monitor and protect data in use (endpoint actions), in motion (network actions) and at rest (storage) — through inspection of content and the context of a transfer, within a centralised management framework.
In practice, a DLP solution:
- finds sensitive data — by patterns (card, passport or tax numbers), keywords, sensitivity labels;
- watches channels: email, web services, cloud storage, USB media, printing;
- applies a policy: allow, warn the employee, block, or log the event for investigation.
How DLP helps with AI
- Web channel control. Policies can cover web services, chatbots included: warn or block when sensitive data is pasted or uploaded into a form. The exact capabilities depend on the product.
- Separating services. A corporate AI service with a contract and configured protection is allowed; personal accounts and unknown services are restricted.
- Logs. If a leak does happen, you can see who sent what and where.
Not just DLP: rules for using AI
Technical control works when there are clear rules alongside it:
- A list of allowed services — preferably corporate versions with a contract that sets out how data is processed.
- What must not be entered: personal data, client data, trade secrets, passwords and keys, unpublished financial results.
- Data classification: which data is public, internal or confidential — so that both DLP and people know what to protect.
- Tidy access rights before rolling out corporate AI assistants.
- Checking the answers. Microsoft warns directly that generative AI responses aren’t guaranteed to be 100% factual, so a person reviews the result.
- Training. Among the measures, OWASP names user awareness: people should know how to work with such services safely.
How to roll out DLP
- Decide what to protect: which data is most sensitive and where it is stored.
- Start in monitoring mode: first only log events, to see real data flows without stopping work.
- Introduce policies gradually: from warnings to blocking.
- Allow exceptions for legitimate business processes.
- Explain to employees why this is done and what to do if an action is blocked.
- Review events and adjust the rules.
Checklist
- Rules for using AI and a list of allowed services are approved
- Employees know which data must not be entered into chatbots
- Data is classified, and confidential data is labelled
- Access rights were reviewed before rolling out corporate AI assistants
- DLP covers email, web services, cloud and USB media
- DLP policies were introduced gradually, and events are reviewed regularly
- AI answers are checked by a person
How Bereg-IT can help
We implement data loss prevention with Forcepoint DLP: we identify sensitive data and channels, set up policies, start in monitoring mode and move to blocking step by step. At a consultation or audit we help with data classification, access rights and rules for using AI.
The first consultation is free. Send a request and an information security specialist will get in touch.
Sources
- OWASP Gen AI Security Project, LLM02:2025 Sensitive Information Disclosure: https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/
- NIST CSRC Glossary, data loss prevention (definition from CNSSI 4009-2022): https://csrc.nist.gov/glossary/term/data_loss_prevention
- Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-privacy