Please report security problems privately through GitHub's private vulnerability reporting: on https://github.com/Inferensys/bulwark, open the Security tab and choose Report a vulnerability. Do not open a public issue, pull request or discussion for a security problem.

Include what you found, the affected files or versions, steps to reproduce, and the impact you expect. We will acknowledge the report, keep you informed while we investigate, and credit you in the fix unless you prefer not to be named.

In scope:

• the Python package (bulwark/), its command-line interface and the scripts in scripts/;
• the job and cloud scripts in infra/ (for example anything that could leak credentials, storage access tokens or data, or leave cloud resources running);
• released model files and checkpoints.

Out of scope: vulnerabilities in third-party dependencies or upstream models, which should go to their maintainers (tell us too if Bulwark is affected), and questions about model accuracy or misuse, which belong in a normal issue.

Bulwark is early research code. Fixes go to the main branch and the next release; older versions are not patched.

• PyTorch checkpoints (.pt) are Python pickles. Bulwark reads them with torch.load(weights_only=True), which restricts unpickling to tensors and plain Python types, and requires PyTorch 2.6 or later, which fixed a known bypass of that restriction (CVE-2025-32434). That narrows the risk without removing it: load checkpoints only from sources you trust. The official model repositories hold no .pt files: https://huggingface.co/Inferensys/bulwark-b1-driver-monitoring holds a B1 release directory, whose weights are in safetensors (which stores no code), and https://huggingface.co/Inferensys/bulwark-e1-driver-monitoring-onnx holds E1's ONNX graphs. Treat any .pt file said to come from an official repository as untrusted.
• Verify the checksums published with a release (SHA256SUMS in the repository; shasum -a 256 -c SHA256SUMS) before using its weights or ONNX files.
• Training data and model inputs are video of people. Treat them as personal data: restrict access, and do not attach them to issues or reports.