Meta has published the engineering behind a promise it made at Connect: that the AI running on its glasses should not be able to hand your data back to Meta itself. The system, called Private Processing, runs AI models inside confidential virtual machines hosted on trusted execution environments, chip hardware that encrypts a machine’s memory under a key never released to the operating system, the hypervisor or the staff running the servers.

Before anything leaves the frames, the glasses demand a remote attestation from Meta’s servers and cross-check the binary fingerprints against a public transparency ledger. If the check fails, the device refuses to connect. Requests travel through an outside relay, and long-term memories are encrypted with keys the wearer holds before they leave the sealed chip. Meta stores only ciphertext.

The company says it has widened its bug bounty to cover Private Processing and is having the design audited by outside firms including NCC Group, so the claims can be tested rather than simply trusted.

A simpler control is arriving alongside it. A toggle tucked into the Meta AI app now lets a wearer keep the imagery its assistant captures out of Meta’s model-training pipeline. The company’s own documentation concedes that such training can be handled by software or by people. Switching it off can cost you features such as visual reminders, and anyone owning more than one pair has to change the setting on each.

Source: Engineering at Meta, Engadget