LocalFlow is a small, offline Windows app maintained by Datafying Tech. We take reports seriously and will answer every one, even the ones that turn out to be nothing.

Please do not open a public issue for a security problem.

Use GitHub's private reporting instead:

1. Go to the Security tab.
2. Click Report a vulnerability.
3. Tell us what you found, how to reproduce it, and what an attacker could do with it.

If that form is not available to you, email kevin@datafying.tech with LocalFlow security in the subject line.

Please include:

• LocalFlow version (.\.venv\Scripts\python.exe -m localflow --version)
• Windows version, and whether you run LocalFlow as administrator
• Steps to reproduce, and a proof of concept if you have one
• Your --doctor output if it is relevant. It contains no dictated text, but it does list your audio device names, so check them before posting.

What to expect. We aim to acknowledge a report within 3 business days and to have a fix or a clear plan within 30 days. We will credit you in the release notes unless you would rather we did not. This is a free project with no bug bounty, but we will not be difficult about disclosure: tell us first, give us a reasonable window, and publish whatever you like afterwards.

In scope:

• Code execution, privilege escalation, or file overwrite triggered by a crafted config.yaml, history.jsonl, or a model file in models/
• Anything that makes LocalFlow send audio, transcripts, or telemetry off the machine
• Clipboard or keystroke handling that leaks dictated text to another process beyond the normal paste into the focused window
• Writing dictated text somewhere the user was not told about (anything outside history.jsonl and localflow.log)
• Dependency vulnerabilities that are actually reachable from LocalFlow's code paths
• Anything in install.ps1 that could be hijacked to run untrusted code during install

Out of scope:

• Attacks that require an attacker who already has code execution on the machine as your user. Such an attacker can already read your clipboard and your keystrokes; LocalFlow cannot defend against that and does not try to.
• The fact that history.jsonl is plain text on disk. That is documented, on purpose, and you can turn it off with history.enabled: false.
• Antivirus or SmartScreen warnings about the global hotkey listener or synthetic keystrokes. That behaviour is the app working as designed and is explained in the README.
• Anti-cheat systems reacting to synthetic keystrokes.
• Vulnerabilities in Ollama, Python, or the NVIDIA driver. Report those upstream.
• Reports produced only by an automated scanner, with no working path through LocalFlow's code.

LocalFlow takes no network input. It opens no listening socket, has no server, no update check, and no telemetry. The only outbound requests it ever makes are to http://127.0.0.1:11434 (your own Ollama, on loopback) and, exactly once, to Hugging Face to download the speech model if it is not already in models/. After that the app sets HF_HUB_OFFLINE=1 on itself and runs with no network at all.

That means there is no remote attack surface to speak of. The realistic threats are local: a malicious config.yaml, a tampered model file, or a supply-chain problem in a pinned dependency. Those are the ones we care most about hearing.