Use GitHub's private vulnerability reporting on the repository Security tab. If it is unavailable, contact the maintainer through a private channel rather than opening a public issue. Include affected versions, impact, and minimal reproduction steps.

Treat model repositories, checkpoints, and cached artifacts as untrusted inputs unless you control their source. SAFT does not enable Hugging Face remote model code by default and uses tensor-only loading for its custom PyTorch artifacts. Transformers and PEFT may still read standard legacy .bin checkpoints; prefer safetensors and load checkpoints only from trusted publishers. Enable any explicit remote-code option only for a source you have reviewed.