Skip to content

Security: LwkMoon/FlowCommand

SECURITY.md

Security policy

Supported versions

Security fixes are made to the latest release. Please update before reporting.

Reporting a vulnerability

Please do not open a public issue for a security problem. Use GitHub's private reporting instead: Security tab → Report a vulnerability on this repository. If that option is unavailable, contact the maintainer privately through the contact details on their GitHub profile.

Include the FlowCommand version, Windows version, what you did, and what you expected. You can expect an acknowledgement, and a fix or an explanation as soon as the maintainer is able. This is a volunteer project, so there is no guaranteed timeline.

What counts

Examples of in-scope problems:

  • A way to make FlowCommand run something outside its action catalogue (for example through crafted AI output, a crafted dictated phrase, or a crafted config.json field that should have been rejected).
  • Bypassing validation, the enabled-action allowlist, risk levels or required confirmations.
  • Anything in the offline interpreter or the welcome and settings screens that lets text become an action outside the catalogue.
  • Exposure of API keys (in logs, the config file, network requests to anyone but the chosen provider, or on disk unencrypted).
  • The key listener recording, storing or transmitting anything other than "was this the trigger key".
  • Closing or terminating protected system processes.

Out of scope: problems that need an attacker who already controls your Windows account or your config.json, the behaviour of the AI providers themselves, and findings in third-party software.

How FlowCommand is designed to limit risk

  • The AI only chooses from a fixed action list; its answer is validated before anything runs.
  • There is no shell, script, arbitrary-file or registry action.
  • URLs and file paths come from configuration, never from the AI.
  • API keys are encrypted with Windows DPAPI for the current user and are not written to logs or config.json.
  • The only network traffic is the request to the AI provider you configured, and only when you choose an online provider. The default offline mode sends nothing. A test keeps the privacy policy in step with the addresses in the code.
  • The keyboard listener checks one key and cancels on others. It can be paused from the tray (no hook or hotkey is registered while paused).
  • Microsoft Store builds run as a normal user (full-trust desktop app, no administrator rights) with no microphone or camera capability. Windows signs the Store package; zip builds are not signed.
  • The offline interpreter produces the same structured plan as an AI would, and that plan goes through the same strict validation and risk policy, so it cannot run anything outside the catalogue either.

There aren't any published security advisories