Security

Runway only calls your bank’s read endpoints: accounts and transactions. There is no code path that pays, transfers or cancels anything.

The agents (Copilot and Autopilot) can only create drafts. A database rule rejects any action an agent tries to create in any state other than draft, and nothing is sent until a person approves it.

Burn, runway, forecasts and savings are computed by deterministic code, in whole cents, from your transactions. Language models only label unfamiliar vendors, answer questions using those figures and the transactions they look up, and write drafts. The privacy page lists exactly what is sent. Every finding links to the transactions behind it.

• HTTPS only, with HSTS. The database connection uses TLS with certificate verification.
• Data lives in Postgres with row-level security on every table; the early-access list can’t be read by any public or signed-in role.
• A content security policy that loads nothing from other sites, and the app can’t be framed.
• Writes from other websites are refused, and the signup and agent endpoints are rate-limited.

The grader runs entirely in your browser. Your file is never uploaded: an automated test checks that the grading code uses no network APIs, and a browser test saw no requests while a file was graded.

Runway is in early access. Today each deployment serves one company behind a password; per-user sign-in and team permissions come before general availability. Rho connections use a company API token until partner sign-in with Rho is available.

Please report security problems privately through GitHub’s private vulnerability reporting, not a public issue. Send a private message through GitHub’s private reporting form (only the maintainer can read it). Please don’t post personal details in a public issue.