Castwright is a local-first, single-user application: by default the analyzer and the voice engine run on your own machine and nothing leaves it. The remote attack surface is therefore small and almost entirely opt-in — it appears only when you deliberately expose the app to your home network (npm run start:lan) or pair the Android companion. The 2026-05-31 security review documents that surface and the hardening backlog.

Security fixes land on main and ship in the next tagged release. Only the latest published release is supported — please reproduce on the newest version before reporting.

Please report privately — do not open a public issue for a security bug.

1. Preferred: use GitHub's private vulnerability reporting — the "Report a vulnerability" button under this repository's Security tab (Security → Advisories). This keeps the report confidential and threads the fix through a private advisory.
2. Alternative: email hello@castwright.ai with SECURITY in the subject. PGP available on request.

Please include: affected version, environment (OS, GPU, local vs. LAN/companion exposure), reproduction steps, and impact. A proof of concept helps but isn't required.

• Acknowledgement within 5 business days.
• An initial assessment (severity + whether it's in scope) within 10 business days.
• We'll keep you updated through the fix and credit you in the advisory unless you'd prefer to remain anonymous.

This is a solo-maintained project — timelines are best-effort, not contractual. There is no paid bug-bounty program.

In scope — issues that let an attacker who is not the local user do something they shouldn't:

• Authentication / cert-pinning bypass on the LAN exposure surface (start:lan) or the companion pairing flow.
• Remote code execution, path traversal, or arbitrary file read/write reachable over the network.
• Code-execution from loading an untrusted shared voice artifact (see side-13 in docs/BACKLOG.md).
• Secret leakage (API keys, tokens) from the running server.

Out of scope (by design, not vulnerabilities):

• The opt-in cloud analyzer (Gemini) sending your manuscript text to Google — that is the documented trade-off of choosing a cloud analyzer; the local Ollama analyzer keeps everything on-device.
• Anything requiring local access to a single-user machine the attacker already controls (the app trusts the local user by design).
• Model output quality / mispronunciation (use the in-app issue forms).
• Denial of service against your own local instance.

Thank you for helping keep Castwright users safe.