Use GitHub's private advisory form: https://github.com/manuc66/TranscribeCppSharp/security/advisories/new.

Please do not open a public issue for a vulnerability. Include the package and version (TranscribeCppSharp, TranscribeCppSharp.Native.<rid>, the CLI or the server), the platform, and the smallest reproduction you have.

This repository is a packaging and binding layer over transcribe.cpp. In scope: the C# wrapper, the generated interop bindings, the CLI, the desktop app, the experimental HTTP server, the packaging and the release workflow.

Not this project's code: the native library and ggml. A crash or a wrong result inside decoding or inference belongs upstream; it will be relayed and the pin in build/TRANSCRIBE_VERSION bumped, but the fix is theirs.

Not sandboxed: model files. They come from the user or from HuggingFace (pinned revision, sha256 verified before the file enters the cache), and they are parsed by the native library in this process. A malicious GGUF is a native-code trust decision, not something the wrapper can contain.

• Native release archives are pinned by sha256 (build/native-sha256.json) and the installed files are re-hashed before tests and packing (build/native-libs-sha256.json, FetchNative --verify).
• Model downloads are pinned to a revision and verified against the LFS sha256 before they are moved into the cache.
• The workflows pin every action by commit sha, default to contents: read, and run gitleaks over the full history.

There is no bug bounty and no third-party audit. Nothing here is claimed to be thread-safe, memory-safe or leak-free beyond what the tests and the ABI checks establish. If something is unknown, it is written down as unknown.