Boundaries
| Boundary | Claim | Why it holds |
|---|---|---|
| NO PRIVATE KEYS | NOSHASHI never asks for, generates or stores an XRPL secret, seed or private key. | No XRPL wallet or key-pair library is imported. The app will not send a message containing an XRPL secret seed to any AI model: it checks every message before sending and refuses. The keys NOSHASHI does hold are of other kinds: model-provider API keys and the Compliance API secret sit in the operating system keyring, and nsh_live_ API keys are stored only as SHA-256 hashes. |
| NO CUSTODY | NOSHASHI holds no client assets and controls no account on the ledger. | Addresses are entered to be read. There is no deposit address, balance under management or withdrawal path. |
| NO SIGNING | No code path signs an XRPL transaction. | The signing commands (sign, sign_for, submit_multisigned, wallet_propose, channel_authorize) appear nowhere in the covered code, which a test checks. NOSHASHI does sign messages of its own kind — webhook deliveries carry an HMAC — which is unrelated to ledger transactions. |
| NO BROADCAST | Nothing is ever submitted to the ledger. | The submit command appears nowhere in the covered code, which a test checks. Every command that is sent is on the published read-only list. |
Where data goes
| Who | What | Why |
|---|---|---|
| Public XRPL servers | The addresses, transactions and order books you ask about. On the website's home page, a subscription to the public ledger stream that drives the pond, which shows the server your connection but no address or query. | To read the ledger. These servers are run by third parties and see your queries. |
| Your device | The workstation's verdict ledger, settings and cases outside an organization. | Kept locally on purpose; not transmitted. |
| Supabase (NOSHASHI's hosted database) | Account, subscription, organization data, shared cases, audit log, API receipts, hashed API keys. | Accounts, billing entitlements and anything shared across an organization. |
| Stripe | Payment details, entered on Stripe's own page. | Billing. NOSHASHI never sees card numbers. |
| Have I Been Pwned | The first five characters of a SHA-1 hash of a new password, sent from NOSHASHI's server. | To refuse passwords known from breaches. The breach service receives only that hash prefix; authentication passwords are handled by the hosted Supabase Auth service. |
| An AI provider you choose | Your questions and the context attached to them. | Only if you configure a hosted provider. The default is a model on your own machine, and the app shows which applies. |
What is not claimed
- NOSHASHI trusts the public servers it reads from. It does not verify ledger proofs or validator signatures itself, and individual reads are not cross-checked between servers (the Network view compares servers separately).
- A receipt digest shows the receipt was not altered. It is not a signature and does not by itself prove NOSHASHI issued it.
- If you configure a hosted AI provider, your prompts are processed by that provider under its terms.
- NOSHASHI holds no security certification (such as SOC 2 or ISO 27001) and makes no claim of regulatory compliance. It is software that reads the ledger and applies rules; the decisions stay yours.
The same model, with each stage's source files, is on the Trust & security page.
How NOSHASHI is built to resist compromise, and — just as importantly — what is not yet hardened. Report anything you find to security@noshashi.app; we ask for coordinated disclosure and reply within one business day.
The core principle
The safest data is data we do not hold. Without an account the app has no server-side state at all: it reads public ledger data, keeps preferences in a local file, and never identifies you to anything we operate. Signing in adds exactly the data listed in the privacy policy and nothing more.
Desktop application
| Control | Implementation |
|---|---|
| Content Security Policy | Strict allow-list. default-src 'self', no unsafe-eval, object-src 'none', frame-src 'none', frame-ancestors 'none'. Every outbound host is named explicitly. |
| Global Tauri object | withGlobalTauri: false — the web view has no ambient bridge to invoke commands. |
| Capabilities | Minimal permission set. The window may show/hide/minimise itself and use notifications, store, autostart, shortcut and positioner. Nothing else. |
| Filesystem | No filesystem plugin. The only write path is one command that refuses path separators, resolves inside ~/Downloads, and caps output at 64 MB. |
| Shell | No shell plugin. Opening a URL goes through one command that parses the URL and refuses any scheme other than http/https. |
| Secrets | Held in the macOS Keychain via the keyring crate, never in a preferences file, browser storage, or a log. The compliance API secret is write-only from the UI and is never read back into the web view. Model-provider keys are readable by the front end at call time — the request to the provider is made from the web view, so the key has to reach it — and are scoped one keyring entry per provider so revoking one does not disturb another. Routing those calls through the native layer, so the key never enters the web view at all, is tracked as a known gap below. |
| Fonts | Self-hosted. No font CDN is contacted, so opening the app announces itself to nobody and it works fully offline. |
| Telemetry | None. No analytics, no crash reporting, no advertising identifiers. |
Authentication
- Passwords are passed straight to Supabase Auth, which hashes them with bcrypt server-side. The application never stores, logs, or transforms a password.
- Minimum twelve characters with mixed case, a digit and a symbol, enforced before the value leaves the field, with a strength meter.
- TOTP two-factor (RFC 6238) via authenticator app, and email one-time codes as an alternative sign-in path.
- Sessions use PKCE and rotating refresh tokens.
Data isolation
- Row level security is enabled on every table. Each policy scopes rows to
auth.uid(), so one account can never read another's. entitlementsandsubscriptionsare read-only to clients. They are written exclusively by the Stripe webhook using the service role, so a tampered client cannot grant itself a paid plan.- The service-role key exists only inside Edge Functions and is never shipped to any client.
Payments
- No card detail ever reaches this application. Checkout and card management both happen on Stripe's own domain.
- The checkout function accepts only price IDs on a server-side allow-list, so a modified client cannot invent a cheaper plan.
- Seat quantities are clamped server-side (1–500).
- Webhook authenticity is established by verifying the
Stripe-SignatureHMAC in constant time, with a five-minute timestamp window that makes a captured request useless afterwards.
API keys
- Generated from the platform CSPRNG (256 bits).
- Stored only as a SHA-256 digest. Losing the database leaks no working key, and support genuinely cannot recover one.
- Displayed exactly once, at creation.
The AI agent
- Defaults to a local runtime. Prompts containing wallet addresses, receipts and rule traces never cross the network boundary.
- Remote providers are opt-in, restricted to a named allow-list in the CSP, and must be HTTPS — the endpoint validator refuses plaintext for anything that is not loopback.
- The interface states plainly, on the screen where you choose, whether the selected runtime is local or remote.
- The system prompt forbids the agent from accepting seed phrases, keys or passwords, and from issuing verdicts itself.
Known gaps
Named rather than hidden. Last reviewed 2026-09-29 against the code.
- Builds are not code-signed or notarised. macOS Gatekeeper and Windows SmartScreen warn on first launch. Signing needs the company's Apple Developer ID (with notarisation) and a Windows code-signing certificate;
docs/RELEASE_SIGNING.mdlists exactly what to add. Until then each release publishes SHA-256 checksums and updater signatures, and the app hashes its own binary (Settings › Binary integrity). style-srcstill allows'unsafe-inline'. Tailwind and the animation layer write inline style attributes. Removing this needs a nonce-based build pipeline. Scripts are already'self'only.- Supabase Auth's own leaked-password toggle is off. It is a paid-plan setting. Passwords are screened against known breaches by NOSHASHI's own hook instead (
noshashi-password); a breached password is refused. - No independent penetration test or certification (SOC 2, ISO 27001). Controls are documented in
docs/SECURITY_THREAT_MODEL.md; do a penetration test before taking institutional production traffic. - Live posture indicators are not gated on source agreement. Recorded verdicts are (SOURCE_AGREEMENT); the tray, Mission Control and status-rail indicators evaluate the rules without it. They carry no receipt.
Closed since the last review: Edge Functions and the Compliance API are rate limited per key (durable windows in noshashi.api_rate_windows); the Compliance API (noshashi-verify) is deployed; hosted-model API keys are held in the OS keyring and used only by the Rust layer (model_request), never passed through the web view; the web view can no longer toggle developer tools (core:webview:allow-internal-toggle-devtools removed).
Generated from src/lib/trust/boundary.json · SECURITY.md