3.1.2. Data Loss Prevention (DLP)
💡 First Principle: DLP policies watch for sensitive information being shared in ways that violate your rules — and then take action: warn the user, require justification, or block the action entirely. DLP doesn't prevent users from creating sensitive content; it prevents inappropriate sharing of that content.
Without DLP, a user can accidentally paste 50 social security numbers into a Teams chat, email a spreadsheet of customer credit card data to a personal email, or upload a document containing healthcare records to a public SharePoint site — and nothing stops it. DLP is the automated policy enforcement layer that catches these in real time.
A DLP policy has three parts:
1. Sensitive information types (what to look for):
- Built-in types: credit card numbers, SSNs, passport numbers, IBAN codes, health record IDs
- Custom types: patterns you define (e.g., your internal employee ID format)
- Trainable classifiers: ML models, trained on sample items, that recognize content by what it is ("financial reports", "source code") where no fixed pattern exists
- Each sensitive-information-type condition can set an instance count — the minimum number of matches in one item before the rule triggers. Items with fewer matches aren't caught, so the threshold trades noise against coverage
2. Locations (where to watch):
- Exchange Online (email)
- SharePoint and OneDrive (files)
- Teams chat and channel messages
- Endpoint devices (files being copied to USB or printed)
- Microsoft Copilot (formerly Microsoft 365 Copilot) interactions
3. Actions (what to do when a match is found):
- Notify the user with a policy tip (visible in the app, e.g., "This email may contain sensitive information")
- Require override justification (user can override with a business reason, which is logged)
- Block the action (prevent sending/sharing)
- Alert administrators in the Defender or Purview portal
Deploy and tune safely: run a new policy in simulation mode first (it replaced the older test mode): the policy runs as if enforced but blocks nothing, and its matches appear in a separate simulation dashboard. Review the matches, separate true positives from false positives (content that only looks like the sensitive type, such as dummy or sample values), and tune before enforcing — adjust the instance-count threshold or confidence level, add exceptions for the legitimate pattern, and allow a business-justification override for borderline cases. Tuning keeps the protection; switching the policy off or carving whole groups or locations out of its scope removes it. A false negative is the opposite problem: real sensitive content that wasn't detected.
⚠️ Exam Trap: DLP policies operate inside the organization too — not just on external sharing. A DLP policy can block sharing of credit card data in a Teams chat between two internal employees. "Data Loss Prevention" implies only outbound protection, but the scope is broader.
Reflection Question: A financial services firm wants to prevent employees from sending emails containing account numbers to external recipients, but allow internal sharing with a warning. How many DLP policies are needed, and what are the key differences between them?