Every AI assistant's privacy policy says some version of "trust us." AdaShield exists to replace that sentence with something you can actually check yourself — built into the same CLI infrastructure, OS kernel boundary, and mail pipeline AdaDo already runs on, instead of bolted on as a separate product.
AdaDo already separates what's PII from what isn't at an architectural level (AIRLOCK/AIRGAP). AdaShield's job is making that boundary independently checkable by you — or a third-party auditor — instead of something you just have to believe we honour internally.
AdaDo's agents already run through real CLI infrastructure (Claude Code, Codex). That gives us a real enforcement point most assistants don't have: per-session audit trails, credential-scope limits, and anomaly detection on exactly what a session touched.
AdaOS's research architecture mediates syscalls at the init boundary rather than replacing the kernel. That boundary is a natural place to enforce security decisions at the device level, not just inside one app.
Spam, phishing, and malware filtering built into AdaDo Mail's pipeline directly — not a bolted-on scanner running after the fact.
Define and publish exactly how a user or auditor can independently check the AIRLOCK/AIRGAP boundary is being honoured — the trust anchor, built first because it depends on nothing else being finished.
Real per-session logs of credential scope and data touched, surfaced somewhere a user can actually see them — not buried in an internal log only Ada reads.
Filtering built into AdaDo Mail's real pipeline as soon as it reaches phase 1.
Device-level security decisions at the syscall/init boundary, once the underlying AdaOS research is far enough along to build on.