Whisper Face controls a microphone, reads bounded application context, and can insert text into another application. Security and privacy failures in those boundaries are treated as product failures, not merely bugs.

Security fixes are made on the main branch and the newest published release. Older builds may not receive backports. The exact source revision for a running build is available from Diagnostics -> Exact Source and from GET /source.

Do not open a public issue for an unpatched vulnerability or include another person's audio, transcript, credentials, private context, or recovery text in a report.

Use GitHub's private vulnerability-reporting form:

https://github.com/Aiml3ss/whisper-face/security/advisories/new

Include the affected revision, macOS or Windows version, reproduction steps, impact, and the least-sensitive evidence that demonstrates the problem. Please say whether the issue is known to be actively exploited.

The project will try to acknowledge a complete report within three business days, provide an initial severity assessment within seven business days, and coordinate a fix and disclosure date with the reporter. These are response targets rather than warranties. A critical, reproducible issue may require an immediate release or temporary feature disablement.

Good-faith research should use accounts, machines, speech, and documents you own or have explicit permission to test. Stop if testing could expose another person's data or disrupt a service. Do not use social engineering, persistence, destructive payloads, denial of service, or public disclosure before a fix is available. The project will not pursue action for research that follows these guidelines, avoids privacy harm, and is reported promptly, to the extent the Project Owner can make that commitment.

• capture continuing after the UI says it stopped;
• audio, transcript, context, correction, or clipboard data leaving the device unexpectedly;
• unauthenticated access to the transcription endpoint outside its intended trusted network;
• focus/selection confusion, duplicate insertion, or insertion into the wrong application;
• update, installer, model, or dependency supply-chain compromise;
• release checksum, signature, notarization, or source-revision mismatch;
• local private files created with permissions broader than the current user.

The public trust boundaries and current mitigations are documented in the threat model. What a downloaded release does and does not prove about its own origin is documented in release provenance. The user-facing data contract is the privacy promise.