RLX is pre-1.0. Security fixes land on the latest 0.2.x release; there is no backport guarantee for older versions. Pin an exact version in production and upgrade to pick up fixes.

Please report suspected vulnerabilities privately — do not open a public issue for anything exploitable.

• Preferred: open a private security advisory via GitHub (Security → Report a vulnerability) on MIT-RLX/rlx.

Include enough to reproduce: affected version/commit, platform and backend (CPU / Metal / CUDA / …), a minimal graph or input, and the observed vs. expected behavior. We aim to acknowledge within a few business days and will coordinate a fix and disclosure timeline with you.

RLX compiles and executes model graphs and loads model weight files (GGUF, safetensors, ONNX). Reports involving memory-safety in kernels or loaders, malformed-model-file handling, or unsafe FFI boundaries are in scope. Running untrusted models or kernels is inherently sensitive — treat model files from unknown sources as untrusted input.