The DID password, wallet signature, address, fula peer id and root CID (the values the mobile app keeps in the OS keychain) are stored in IndexedDB (fxblox-secure) encrypted with AES-GCM-256 under a non-extractable WebCrypto CryptoKey that itself lives in IndexedDB. Each record carries AAD fxblox|secure|v1|<service> so a ciphertext cannot be moved between slots. wipe() deletes the records, the key and the database (logout / "clear cached data").

Trust model, stated plainly:

• Protects the secrets at rest and prevents key exfiltration (the key cannot be exported by script).
• Does not protect against script running on this origin: an XSS could use the key while the page runs. The Content-Security-Policy (no third-party scripts, script-src 'self' 'wasm-unsafe-eval') is therefore the primary control; the design assumes CSP is not bypassed and the browser's crypto isolation holds.
• Identity is cheaply re-derivable from the password + one wallet signature, so loss of the store is a 30-second "re-link", not data loss.
• Deferred hardening: PIN-wrapped unlock and idle auto-lock.

• :3500 / :8083 are plain HTTP on the LAN/hotspot. Chrome's Local Network Access permission gates the browser side; firmware PRs add a CORS allow-list (https://blox.fx.land, https://functionland.github.io, http://localhost:*) and an Origin guard so a cross-site simple-form POST from another page cannot drive the Blox.
• libp2p actions are authenticated by Ed25519-signed headers bound to the app's peer id (the Blox authorizer); the browser presents the same identity as the phone. A 401 is retried once with a corrected clock, then surfaced as "not authorized".

See apps/fxblox-web/index.html. connect-src allows private-IP http://*:3500 / http://*:8083 hosts (CSP cannot express CIDRs); the client additionally refuses any non-RFC1918 target before issuing a LAN request.