The following versions of Shrew are currently supported with security updates:

I take the security of Shrew seriously. If you discover a security vulnerability, I appreciate your responsible disclosure.

1. Do NOT open a public issue. Security vulnerabilities should not be reported through public GitHub issues.
2. Use GitHub Security Advisories: Open a draft security advisory in the repository.
3. Email: You can also reach me directly at jsimancas@unimagdalena.edu.co.

When reporting a vulnerability, please include:

• A description of the vulnerability and its potential impact.
• Steps to reproduce the issue or a proof-of-concept.
• The affected crate(s) (e.g., shrew-core, shrew-cuda, shrew-ir, shrew-python).
• The affected version(s).
• Any suggested fix or mitigation, if available.

• Acknowledgment: I will acknowledge receipt of your report within 48 hours.
• Assessment: I will provide an initial assessment within 7 days.
• Resolution: I aim to release a fix within 30 days for confirmed vulnerabilities, depending on complexity.

• You will be kept informed about the progress of the fix.
• If the vulnerability is confirmed, I will issue a patch release and publish a security advisory.
• If the report is declined, I will provide an explanation.
• I will credit reporters in the security advisory (unless you prefer to remain anonymous).

Shrew uses unsafe Rust in specific, well-scoped areas:

• shrew-cuda: FFI calls to CUDA driver/runtime (cuBLAS, custom PTX kernels) via cudarc.
• shrew-cpu: SIMD-accelerated operations through gemm.
• shrew-python: PyO3 FFI boundary for Python interop.

All unsafe blocks are documented and minimized.

• .shrew checkpoints, Safetensors, and ONNX model files involve deserialization. Only load models from trusted sources.
• The .sw IR parser validates input before execution, but .sw files from untrusted sources should be treated with caution.

I regularly audit the dependency tree. Key dependencies:

The shrew-cuda crate executes GPU kernels and manages device memory. It requires a trusted CUDA toolkit installation. Do not use the CUDA backend with untrusted PTX kernel sources.