Skip to content

Security: thc1006/whispersprint

Security

SECURITY.md

Security and privacy

Process media you are allowed to upload. Recordings run on Google's Colab VM; no external ASR API is called. This is not offline processing. Packages/model assets are downloaded from their publishers; hashes validate the pinned model contents, not the trustworthiness of an arbitrary notebook publisher.

Only timestamped TXT is automatically downloaded. Local logs, journals and JSON can include filenames, generated text and tokens. Review/redact selected evidence before sharing; the app does not send it anywhere or auto-download an error archive. Do not publish source recordings, full private logs, credentials or model weights.

The app uses SafeTensors, no remote model Python/pickle execution, local-file FFmpeg protocol restrictions, bounded upload paths and a private JIT cache. Do not import another user's compiled cache. The isolated worker is controlled through stdin, not a public port. Only its own process group may be stopped; there is no quota bypass, keepalive traffic, privilege escalation or sysfs modification.

Report security issues privately to the maintainer using the repository's security reporting route when available. Do not attach sensitive data to public issues.

There aren't any published security advisories