Nepalese Developer

Your glasses can hear the world. Should they obey a TV ad? Meta's own WebMCP docs say the untrusted-content flag is "not an enforced security boundary," and the assistant can't tell your voice from a stranger's. Once the glasses have a camera, a mic and a wallet, a poster, a webpage or a look-alike invoice can give the order. Capability is solved. Authority isn't. I built ONLY YOU to fix that. Your enrolled voice is the only command channel. Scene text, page text and documents are data the agent can read them, never obey them. Anything irreversible needs a signed, scoped, expiring mandate (payee, cap, expiry) that a merchant can check offline with just a public key. Every action and every refusal gets a hash-chained receipt. The demo is a success, not a block. I sign one mandate; a merchant verifies it and gets paid. Then the same invoice arrives with a different payee — refused, because the payee came from a document, not from me. The numbers: a frontier model paid 50% of injected invoices undefended; behind the gate, 0%, with 18/18 receipts. On 12 held-out speakers the voice check accepted 0 impostors at the capped threshold the uncapped pilot let 2 through, which is why the cap is there. 172 tests, CI, Docker. Built on five Meta models Spark, Wonder, Voice Transcribe, SAM 3.1, and Muse Image. What still breaks it cloned voices, the 20% adaptive attacks is published, not hidden. Live: sushant-me.github.io/only-you · verify: /verify.html · repo: github.com/sushant-me/only-you