Trust

Security and trust

Last updated: 12 October 2026

Discovery files often hold a client's plans, names and numbers. This page says plainly how Discovery Phase AI treats them: what is stored, who can see it, what the AI does with it and what you can take back. For personal data and legal terms, the Privacy Policy and the Terms of Service apply.

1. Your files and the AI

Source files (decks, documents, transcripts, images) are read in memory to draft your project and are not stored. They are not written to disk and there is no file table: only the draft you review and save stays, as ordinary project content.

  • When you use an AI feature, the relevant project text and files for that request are sent to our AI provider, Anthropic, to generate the result. We do not use your content to train AI models, and under Anthropic's commercial API terms it does not train on inputs and outputs by default.
  • The AI has no tools. It cannot browse, run code or change your project. Its reply is parsed as data, and every suggestion is a proposal: nothing changes until you apply it, and removing content always asks you to confirm.
  • Text from your files and from other stages is sent to the model marked as untrusted data, so instructions hidden inside a document are not followed. The ids and records an answer refers to are checked against your own project before they are used.
  • AI use is limited per person and per plan, so a runaway script cannot spend a company's credits.

2. Who can see your projects

Each company is its own workspace. The server takes your company from your signed-in session, never from the request, so a project is only reachable by the people of the company it belongs to. We test that one company's list never returns another company's rows.

  • Company Admins see every project of their company. Company Members see only the projects they are added to, as Project Write or Project Read. These checks run on the server for every request, not only in the interface.
  • A Company Admin chooses who joins and with which role. Invitation links work for 7 days and are stored only as a hash.
  • On the Agency plan each client company stays separate: one company never sees another's projects or people.
  • Every change to a stage is recorded in the project history with who made it and where it came from (you, the AI chat, an update from files), and earlier versions can be restored.

3. Accounts and sign-in

  • Passwords are stored only as a bcrypt hash, never in plain text. New passwords must meet a complexity rule.
  • You can sign in with Google. We check Google's signed answer (who issued it, for which app, its expiry and a one-time value) and use PKCE, so a stolen sign-in code is useless on its own. We link a Google account to an existing account only when Google has confirmed the email address, and if that account never confirmed its address, its password and sessions are cleared first.
  • You stay signed in with one cookie that scripts on the page cannot read (httpOnly). In production it is sent over HTTPS only.
  • Only one device is signed in at a time. Signing in somewhere else signs the other device out, so a stolen cookie cannot lock you out of your own account.
  • Failed sign-ins are limited per account and per address, and signing in with an unknown email takes as long as a known one, so the form does not reveal which addresses have accounts.
  • You must confirm your email address before using the app, unless Google has already confirmed it. Resetting a password by email signs you out everywhere.

4. How the application is protected

  • The site is served over HTTPS, with HSTS in production, a strict Content Security Policy, and headers that stop it being framed or content-sniffed.
  • Requests are size-limited and rate-limited, including sign-in, the public forms and every AI route.
  • Rich text is sanitized on the server when it is saved and again in the browser before it is shown. UI screen wireframes are drawn in a sandboxed frame that can run no scripts.
  • Payments are handled by Paddle, our merchant of record. Card numbers never reach our servers. The checkout is created by our server and Paddle's payment notifications are verified by signature, so a plan can only change through a genuine payment event.
  • The production container runs as an unprivileged user and holds no files of yours: all state lives in the database.

We reviewed every route of the application, who may call it and what it can reach, and fixed what we found. This was our own review, not an independent audit.

5. Your data, your control

  • Export: Settings, Personal, Your data downloads a copy of your personal data. Your project content leaves through the discovery report, as PDF, Word, PowerPoint or Markdown.
  • Delete: Settings, Personal lets you delete your account with your password or, for an account without one, by confirming with Google and typing your email. A company you are alone in, with all its projects, is deleted. In a company with other people, what you authored passes to its longest-serving admin.
  • We keep support messages for 24 months, AI usage records (feature, model, token counts; never the text) for 13 months, and product analytics for 13 months, then delete them automatically. The Privacy Policy lists every period.
  • Our own page-view analytics use no cookies and store no IP address, and we honor Do Not Track and Global Privacy Control.

6. Who handles data for us

  • Fly.io hosts the application and its database in the United States (Virginia).
  • Anthropic processes the AI requests described above.
  • Resend sends our email: confirmations, invitations, password resets and a few onboarding tips.
  • Paddle processes payments, taxes and invoices.
  • Google confirms who you are when you choose to sign in with it. We receive your name, email address and a Google account identifier, nothing else.

The Privacy Policy describes what each of them receives.

7. Report a vulnerability

If you think you found a security problem, tell us through the Help and support form (choose "Something is broken" and start the subject with "Security"). Please give us a reasonable time to fix it before you share it, and do not access data that is not yours. We read every report.