If you find a security issue in trtcheck, please do not open a public issue. Use one of:

• GitHub's private vulnerability reporting on this repository. Preferred.
• Email the maintainer at sohams.web@gmail.com with the details.

You should get an acknowledgement within five business days. Fixes typically ship in the next patch release; severe issues get a release out-of-cycle.

trtcheck follows semver from v1.0 onward. Only the current minor receives security fixes:

If you're running 0.x for a reason that prevents upgrading, mention it in your report and we'll discuss backports.

In scope:

• The trtcheck CLI and library, including all bundled checkers, fixers, and reporters.
• The GitHub Action shipped from this repo (action.yml and action/).
• The bundled operator matrix and remediation database.

Out of scope:

• Third-party plugins discovered via the entry-point mechanism. Those are their own packages with their own maintainers.
• TensorRT itself or NVIDIA's trtexec. Forward those to NVIDIA.
• ONNX files supplied by users. trtcheck only reads them; if a malformed ONNX file crashes onnx.load, that's an issue to file against onnx.

• No untrusted code execution. trtcheck loads ONNX files via onnx.load and inspects the resulting graph in memory. It does not deserialize pickle, eval strings, or execute model code.
• HTML reports are self-contained; all dynamic strings from ONNX are HTML-escaped before being written.
• GitHub Action sticky comments escape ONNX-derived text before rendering it as markdown.
• The GitHub Action workflow has strict allowlists on inputs.paths and inputs.version.