Security model

use-everywhere has two very different security postures, matching its two worlds. Knowing which one you're in tells you what the library checks for you and what remains your job — so let's take them in turn, same-origin first because it's the short one.

Same origin: everyone is you

Everything on one origin — every tab, iframe from your domain, worker — runs your code with your cookies. BroadcastChannel is already scoped to the origin by the browser, so the same-origin engines add no authentication at all, on purpose:

• Any code on your origin can join any bus, read state, and write patches.
• A compromised dependency on your page can too — but it could equally read localStorage or your DOM. The bus does not widen that blast radius.

The one same-origin knob is filtering, not security: scope: 'tabs' (or a custom accept predicate on createSharedStore) lets a store ignore writes from workers. That's a coordination tool — a worker is still your code.

Don't put secrets in shared state

State is broadcast to every context on the origin and lives in plain JS memory. Treat it like you treat React state — fine for UI status, wrong for raw card numbers or tokens that deserve tighter handling.

Cross origin: the other page is nobody

The window channel assumes the opposite: the page on the other side is a different security principal, and postMessage is a public mailbox — any window that holds a reference to yours can post to it, and by default a listener cannot tell who sent what. This is the API where real money gets lost, so the paranoia is built in rather than left as an exercise.

Every message a channel receives must pass four gates before your handler ever runs:

Gate	Check	Stops
1. Origin	event.origin === peerOrigin (exact match, no wildcards)	Any site you did not name
2. Envelope	payload carries the library brand (__ue: 1)	Unrelated postMessage traffic on the same page
3. Nonce	cid equals the id minted for this openWindow() call, delivered via the ?ue-cid= URL param	Stale windows, replays from earlier sessions, guessed messages
4. Source	event.source is the exact Window we opened (opener side)	A third frame on the correct origin impersonating the child

Fail any gate, and the message is dropped silently — your handlers never run. Note what gate 4 buys you: even a hostile iframe on the correct origin fails, because it isn't the exact window object you opened.

Outbound is symmetric: every postMessage is sent with an explicit targetOrigin, so the browser itself refuses delivery if the window has meanwhile navigated somewhere unexpected.

peerOrigin is required, '*' throws

openWindow(url, { peerOrigin: 'https://pay.example.com' }); // ✅
openWindow(url, { peerOrigin: '*' }); // ❌ throws

The API is shaped so the insecure thing is not expressible. A wildcard would disable gate 1 and the outbound targetOrigin check at once — that's the textbook postMessage vulnerability, and it doesn't compile here. For local prototyping there is an explicit escape hatch — allowAnyOrigin: true — named so it cannot be mistaken for production configuration. openWindow also cross-checks that the URL you open actually belongs to peerOrigin, catching copy-paste mismatches at call time.

Why shared state stops at the origin line

It's tempting to want useSharedState to just work across the payment window too. The library refuses by design, and the refusal is worth spelling out:

1. Confused deputy. State merging is automatic — whatever arrives with a newer clock wins, no questions asked. Automatic writes from a foreign principal into your application state is precisely how a compromised or buggy partner page would corrupt yours. Messages, by contrast, only do what your explicit, typed handler does. Across trust boundaries, explicit beats magic.
2. Auditability. A typed message map ({ 'payment-complete': Receipt }) is a reviewable contract between two domains. "Any key, any value, newest clock wins" is not.
3. The shapes differ anyway. Cross-origin flows are request/result shaped — open a window, exchange a handful of messages, get one result. A convergent replicated object buys nothing there.

What the library does not protect against

Honesty about the boundary, because trusting a tool means knowing its edges:

• Your own origin. XSS on your page owns every bus. The library neither helps nor hurts.
• The peer's content. Gate 1 verifies where a message came from, not that the payment page is honest. Trust in the counterparty is a business decision; validate the payloads you receive.
• Client-side truth. The demo's payment lock stops accidental double submission. A hostile user can bypass any client-side state — money always needs server-side idempotency keys.

Where to next

• Cross-origin payments — the flow these gates protect, end to end.
• useOpenedWindow — the hook-level view of the same machinery.
• How sync works — the cross-origin handshake in sequence-diagram form.