Skip to content

Security: bzlu-Group/MAD-Operator-Learning

Security

SECURITY.md

Checkpoint and runtime security

Supported public environment

The supported public R4 environment uses Python 3.11 or newer and PyTorch 2.15.0 (requirements-r4.txt). Install it in a separate environment, not on top of an older CUDA/torchvision/torchaudio installation. The paper's recorded training used PyTorch 2.5.1+cu121; that historical fact remains in the protocol and result records. It is not the recommended installation for new users. Updating the public runtime does not retrospectively change the recorded experiments, nor promise bitwise-identical training on a different runtime.

PyTorch 2.5.1 is affected by CVE-2025-32434 / GHSA-53q9-r3pm-6pq6. PyTorch also reports CVE-2026-24747 / GHSA-63cw-57p8-fm3p in versions through 2.9.1. Both concern maliciously crafted checkpoint files; an alert does not by itself establish that a repository, machine, or experimental result was compromised. The upstream advisories are:

Loading models and resuming training

The public R4 evaluators and plotting entry point use restricted weights_only=True loading on CPU by default. A rejected restricted load is never automatically retried with unrestricted pickle.

Full resume checkpoints contain Python/NumPy random state. Only for files you created or whose source you independently trust, supply --trust-checkpoint to a public R4 command that needs such a checkpoint. train_public.py --resume requires that explicit flag. It enables pickle loading, which can execute code even on a patched PyTorch version. Do not use it to make an unknown downloaded file "work". Checksums detect changes only when the expected checksum itself comes from a trusted source; they are not a substitute for establishing trust.

The frozen train_pinn_balanced.py, earlier R3 programs and original benchmark scripts preserve their historical loading behavior. Do not use them directly with untrusted model files. Prefer the guarded public R4 entry points for the final source-term experiments. No claim is made that every legacy workflow has been revalidated on the new runtime.

These changes address known dependency and loading risks; they do not make arbitrary model files or executable research code safe. Keep the runtime updated, avoid unknown checkpoints, and run reproduction in an isolated environment without sensitive files or credentials.

Reporting

Do not publish credentials or exploit payloads in an issue. Report an upstream PyTorch issue using its security policy. For a repository-specific problem, contact the authors using the contact information in the article.

There aren't any published security advisories