Security

VERSION 1.0 · EFFECTIVE SEPTEMBER 9, 2026 · QUANTGPT LLC, A FLORIDA LIMITED LIABILITY COMPANY

// THE SHORT VERSION

QuantGPT is a research terminal. It runs your criteria against a survivorship-free warehouse and hands the numbers to you or to your AI. It never holds your money, never places an order, and never holds the key to your model. Everything below describes the product as it ships today, not a plan. When a line here changes, the version line at the top changes with it.

• CustodyNone. There is no balance to hold and no wallet to drain.
• OrdersNever placed, modified, or cancelled by QuantGPT. Broker links are read-only.
• PasswordsScrypt-hashed with a salt per account. We cannot read them back.
• API keysShown once, stored as a SHA-256 hash, revocable any time.
• Model keyYours, kept in your browser. It never reaches our server.
• CardsHeld by Stripe. We keep a customer id, a plan, and a status.

// YOUR PASSWORD AND SIGN-IN

Passwords are at least ten characters and are stored as scrypt hashes (cost 16384, block size 8, one lane) with a salt generated for each account. A wrong password costs the server the same time as a right one, so a guess learns nothing from the clock. After a handful of failed tries from one address the next attempt waits, and the wait doubles each time, up to an hour.

Two-step sign-in is yours to turn on from the account page: codes from an authenticator app (TOTP) or a passkey (WebAuthn) tied to your device. Signing in sets one cookie. It is HttpOnly, so no script in the page can read it, SameSite Lax, so another site cannot ride it, and Secure over HTTPS in production. The terminal carries no third-party cookies. Signed-in pages are served with no-store, so nothing you see is cached at the edge for the next visitor.

// API KEYS FOR YOUR AI

Keys that connect your AI to the terminal are shown once when issued. We keep only a SHA-256 hash of each key, so a copy of our database does not contain a usable key. Revoke any key from the account page and it stops working on the next call. Each plan has a key limit and a daily call cap, and every call is counted against the account that owns the key, not the key itself, so rotating keys does not reset the meter.

// THE WAREHOUSE

The engine opens the warehouse read-only and turns off external access, so no query can write to it, read a file from the server, or reach a URL. Memory and thread bounds keep one crafted scan from starving everyone else. Criteria in the terminal are compiled from a fixed factor table, never pasted into SQL. The free SQL your AI can run through the MCP endpoint is gated to one SELECT statement at a time, blocked from the account tables, and runs on the same read-only, bounded connection. Extracts are capped at five thousand rows per call. The warehouse itself lives on infrastructure QuantGPT operates, not on a shared multi-tenant database.

// YOUR MODEL KEY

The desk agent inside the terminal runs on a key you bring: Anthropic, OpenAI, Groq, xAI, or Perplexity. That key is saved in your browser's local storage and every model call goes straight from your browser to the provider. Our server sees only the tool calls the model makes (a backtest, a screen, a grade), never the key and never your conversation. Clear the site's data and the key is gone. QuantGPT holds no house model key of its own, so there is nothing of yours in our provider bills and nothing of ours to leak.

// BROKER CONNECTIONS

When you link a brokerage the adapter is read-only: balances, positions, and history come in, nothing goes out. There is no code path in the product that places, modifies, or cancels an order. Linking uses your broker's own authorization flow, so your brokerage password is never typed into QuantGPT and never stored by it.

// BILLING

Payments run through Stripe. Your card number goes to Stripe and never touches our servers; we store your Stripe customer id, your plan, and your subscription status. Stripe's own security program covers the payment data it holds.

// LIMITS AND ABUSE

Browser sessions carry daily quotas on runs, screens, and funnels; API keys carry the per-plan call caps above. Hitting a limit returns a plain message and resets at midnight UTC. It is never a surprise charge. We keep a usage log under your account (pages, studies, screens, key calls) to run the product and to investigate problems; the privacy page says what that log holds and who can see it.

// WHAT WE DO NOT DO

We do not hold funds. We do not place orders. We do not sell data. We run no ad trackers inside the product. We keep no model key of our own. We never see your brokerage login. We do not read your model conversations, because they never pass through us.

// WHAT IS STILL AHEAD

No independent penetration test has been run yet and QuantGPT holds no SOC 2 report. Both are on the road to the public launch; when either is done it will be listed here with a date, not before. Until then the claims on this page are backed by the code, which is the only thing we can vouch for today.

// REPORTING A PROBLEM

Found something? Write to desk@quantgpt.co with SECURITY in the subject line. Every report gets a reply from a founder, and a fix gets credited here if you want the credit. Please give us a reasonable window to fix before publishing, and please do not test against other operators' accounts or data.

THIS PAGE DESCRIBES THE PRODUCT AS SHIPPED, IN PLAIN LANGUAGE. IT IS NOT A WARRANTY. THE TERMS AND PRIVACY POLICY GOVERN; THE CURRENT VERSION OF EACH LIVES AT ITS URL.