Security fixes go into the latest release on PyPI and the master branch. Older releases are not patched; please upgrade.

Please do not report security problems in public issues, pull requests or discussions.

Report them privately through GitHub: Report a vulnerability (Security tab → Advisories → Report a vulnerability).

Please include:

• the affected version or commit
• steps or a minimal script to reproduce
• the impact you expect (for example code execution, data exposure, data corruption)

We aim to acknowledge reports within a week and will keep you updated until a fix is released. If you accidentally committed a credential to this repository, revoke it with the issuing service first, then tell us through the same channel.

PyHealth is a research library that runs with the permissions of the user calling it. Some inputs are loaded as code-capable Python objects and must come from a source you trust:

• Processed dataset and processor directories. SampleDataset(path), load_processors(path) and the litdata sample caches written by set_task() are stored with Python pickle. Loading a directory someone else prepared can run arbitrary code. Only load caches you created, or that come from a source you would trust to run code on your machine.
• Model checkpoints. PyHealth loads checkpoints with torch.load(..., weights_only=True) where it can, but third-party model files (for example Hugging Face models) should still come from trusted sources.
• Downloaded resources. Some datasets and code mappings are downloaded on request (download=True, medcode). Prefer a trusted network and verify large downloads where checksums are published.

Reports that PyHealth executes code from one of these trusted inputs are expected behaviour, not vulnerabilities, unless PyHealth loads them without the user asking it to.