DLP Policy Lab: stop leaks into AI prompts
Northwind just rolled out an enterprise AI assistant. Staff keep pasting sensitive data into prompts. Write DLP rules that catch the risky prompts without blocking normal work. A colleague already wrote one rule... it's not great.
Your DLP rules
How patterns (regex) work, in 30 seconds
\d= any digit,[A-Za-z0-9]= any letter or digit,{16}= exactly 16 times,{20,}= 20 or more[ -]?= an optional space or dash,\b= edge of a word- Example:
sk-[A-Za-z0-9-]{20,}matches API keys that start with sk- and have 20+ characters after it - Keyword rules: words separated by commas, not case sensitive
- "Min matches" lets you catch bulk data (e.g. 3+ email addresses = a customer list) but not one email
Results
| Press "Test rules". |
Goal
Catch at least 7 of 8 sensitive prompts with at most 1 false alarm. Use redact where the user can still get help without the secret, block for things that must never leave, and warn for grey areas.