Meta's PrivacyLib Enforces Privacy at the Data Access Boundary

• PrivacyLib is a common library running at the data access boundary within hundreds of data and processing system clients at Meta, performing billions of checks per second.
• The architecture separates one-time system integration from pluggable privacy modules, allowing new requirements like Policy Zones or Privacy Probes to be implemented in days instead of the months required by previous bespoke integrations.
• The system supports AI agents by propagating agent context through tool calls to PrivacyLib, enabling runtime blocking of unauthorized data access and logging via the Agent Observability module.
• Engineering challenges are addressed through build-time compilation of requirements, in-process caching to avoid network calls, and isolated deployment via service-specific pipelines to limit failure impact.

Architecture and Scale

• Covers all access paths via typed client libraries and low-level network boundaries, mapping native APIs to five common operations: Read, Write, RPC, Request entry, and Inline execution.
• In-process evaluation provides richer context (caller ID, endpoint, destination) and isolates faults to specific service versions, avoiding shared datastore impacts.
• Telemetry uses Bloom filters to downsample high-load patterns and adjust sampling rates for rare events.

Deployment and Safety

• Annotation changes require code review and simulation to verify data flow impacts before release, with artifacts updated by code generation pipelines.
• Module rollouts use runtime controls to gradually increase enforcement from 0% to 100%, allowing immediate rollback by reducing percentages if valid writes fail.
• Direct RPC paths use encrypted signed tokens verified by ServiceRouter, falling back to coarser payload-based checks if tokens are missing.