Data Access Transparency
End-user-verifiable audit protocol for every privileged access — human admins, AI agents, automated systems, anything with credentials. Submitted to the Linux Foundation Agentic AI Foundation working groups.
Draft ready to read
draft-wang-data-access-transparency-00
Full text of the draft, rendered inline. Synced hourly from server/protocol — no stale copy in this repo.
~12K words · last synced 2026-05-06 11:55 UTC
What the spec covers
- Wire-protocol access capture. How every read and write is captured at the database protocol layer — SQL for Postgres, BSON for Mongo, S3 API for object storage — regardless of whether the caller is a human admin, a cron job, a service account, or an AI agent.
- Caller identity labeling. A canonical format for identifying the principal behind each access: human admin via SSO, service account, automated system, or AI agent (with framework, model, run_id, tool_call_id).
- Chain entry schema. Required and optional fields for capturing each access as a cryptographically chained event — uniform across caller types so an auditor reads human and agent activity in the same log.
- User-verifiable verification flow. WASM-based client-side verification of a per-user chain, tamper-evident across every caller category. The operator cannot rewrite history without the user’s own browser catching it.
- Regulatory conformity mapping. How chain entries satisfy GDPR Article 32, HIPAA audit controls, and EU AI Act Article 12 — including the six-month retention floor from Article 19.
Reference implementation · AGPLv3 · shipping ahead of the formal spec