Google invites the privacy community to review the architecture, protections, security proofs, and verification protocols for transparency and oversight.

Details on availability, supported devices, account recovery, deletion behavior, and enterprise controls are still unresolved and will determine alignment with the described architecture.

The initiative is a cross-team effort involving DeepMind, Platforms & Devices, Core, and Cloud, with acknowledged executive sponsors.

Co-developed by Google DeepMind and the Platforms & Devices, Core, and Cloud teams, with executives named as sponsors.

Encryption and enclave protections do not fully eliminate risks from compromised devices or misused permissions, highlighting ongoing identity and access-control challenges in enterprise AI.

Security governance includes no admin access to plaintext data, containment via confidential VMs, default-deny egress for monitoring, and external audits validating the design.

Google unveils a persistent, private server-side memory feature for Private AI Compute, designed to keep user data in a secure cloud vault while keys stay on the user’s devices.

The new memory layer moves beyond the platform’s prior stateless design by storing per-user, encrypted context in device-derived, per-user databases to protect privacy.

Architecture relies on hardware-enforced secure enclaves, end-to-end encryption, and memory isolation, enabling cross-device, private AI by securely retaining user context over time.

Planned next steps include client-side attestation verification, a co-signed append-only transparency log, broader reproducible builds, and ongoing third-party audits.

Privacy and security researchers are invited to review the updated technical brief and verification protocols, signaling openness to third-party validation.

Trail of Bits conducted a security audit, uncovering 10 findings (including two high-severity), eight fixes applied, with some issues still open.