CVE-2025-62609: MLX NULL pointer dereference in load_gguf

Executive brief

MLX is a machine learning framework for Apple silicon. A vulnerability in its GGUF file loading component allows a malicious model file to crash the application. This could lead to a denial-of-service if a user or automated system attempts to process an untrusted machine learning model.

Technical details

A NULL pointer dereference (CWE-476) exists in `mlx/io/gguf.cpp` within the `extract_tensor_data()` function. The vulnerability occurs because the `weights_data` pointer, provided by the external `gguflib` library, is passed directly to `memcpy` without validation. An attacker can provide a specially crafted GGUF file that results in a NULL pointer for tensor weights, leading to a segmentation fault and application crash. This issue is fixed in version 0.29.4 by adding explicit checks for the `weights_data` pointer.

Affected products

• ml-explore mlx <= 0.29.3

Timeline

• 2025-11-21: disclosed
• 2025-11-21: advisory
• 2025-11-21: patched: Fixed in version 0.29.4

References

• https://api.github.com/users/wickgit
• https://github.com/wickgit
• https://api.github.com/users/wickgit/gists%7B/gist_id%7D
• https://api.github.com/users/wickgit/repos
• https://avatars.githubusercontent.com/u/96148289?v=4
• https://api.github.com/users/wickgit/events%7B/privacy%7D