Meta has extended its Private Processing architecture to AI glasses, proposing a protected cloud environment for features that need larger models or long-term personal context. Basic calls and messages can stay on the device, while workloads such as streaming transcription and contextual search would run in confidential virtual machines.
Before sending personal data, the glasses verify a hardware-signed certificate and compare the cloud workload against an append-only transparency ledger. A failed check stops transmission. Third-party relays operated by Fastly or Cloudflare use Oblivious HTTP to separate a user’s identity from the request, while encrypted storage remains inside a trusted execution environment spanning CPUs and GPUs.
Meta says engineers cannot inspect protected memory or capture model inputs and outputs. Operational monitoring is therefore limited to aggregate signals such as resource use, latency and hardware failures. The company plans to publish virtual-machine binaries and documentation for researchers and is working with NCC Group and other independent security reviewers.
This is an architecture announcement, not a broadly available glasses feature. Meta has not said which functions will use Private Processing first or when deployment will begin. The design can reduce what the cloud operator sees, but always-on wearable sensing still raises questions about bystanders, data collection and whether the implementation matches the promised protections.