BlueprintPage 24
24AI SECURITY
Data-Loss Prevention
Not simply “is this data sensitive?” but “is this agent authorized to move this data to this destination for this task?”
- Priority
- P1
- Phase
- Phase 2
- Primary owner
- Security / Data
- Status
- Blueprint
Objective
DLP for agents must operate at the point where data crosses tool, model and destination boundaries. It should understand which agent is moving what data, to which destination, under whose authority and for what business purpose.
What to build
- Classify PII, credentials, payment data, financial data, source code and customer-defined sensitive patterns.
- Inspect prompts, tool inputs/outputs, file transfers and external destinations.
- Apply redact, tokenize, mask, block or approval policies.
- Detect unusual bulk export relative to the agent’s historical profile.
Implementation decisions
- Support structured-field scanning plus unstructured detection.
- Allow tenant-defined dictionaries, regex patterns and data labels from existing DLP systems.
- Preserve hashes/metadata for blocked sensitive content without storing raw values unnecessarily.
- Integrate with Microsoft Purview or other enterprise classifications where feasible.
Definition of done
- A protected sensitive field is redacted or blocked before leaving the authorized boundary.
- Every DLP decision identifies source, destination, agent and matched policy.
- Customers can tune detection without disabling entire control families.
Success metrics
- Sensitive events detected
- Exfiltration attempts blocked
- False-positive rate
- Data classes covered
KEY QUESTION
Not simply “is this data sensitive?” but “is this agent authorized to move this data to this destination for this task?”