CVE-2026-69112: Hugging Face Accelerate path traversal and denial of service in checkpoint loading

Executive brief

Hugging Face Accelerate is a library used to simplify distributed machine learning training by managing model checkpoints and distributed loading. The vulnerability allows attackers to craft malicious checkpoint index files that can read arbitrary files from the system or cause the application to hang indefinitely, disrupting service and exposing sensitive data like model weights or configuration files.

Technical details

The vulnerability is a path traversal (CWE-22) flaw in Hugging Face Accelerate's checkpoint loading routines. The vulnerable functions load_checkpoint_in_model and load_checkpoint_and_dispatch do not properly validate or sanitize the weight_map entries from sharded checkpoint index files. An attacker can craft a malicious checkpoint index containing relative path traversal sequences (e.g., ../../../etc/passwd) or absolute paths, enabling arbitrary file read access. Additionally, attackers can point shard entries to named pipes or other blocking resources, causing the application to hang indefinitely and resulting in denial of service. The attack requires providing a malicious checkpoint index file—typically during model loading—but does not require elevated privileges or network access. No patch has been released as of the advisory date.

Affected products

• Hugging Face Accelerate <= 1.14.0

Timeline

• 2026-08-10: disclosed
• 2026-08-10: advisory

References

• https://api.github.com/users/KMChris
• https://github.com/KMChris
• https://api.github.com/users/KMChris/gists%7B/gist_id%7D
• https://api.github.com/users/KMChris/repos
• https://avatars.githubusercontent.com/u/17026216?v=4
• https://api.github.com/users/KMChris/events%7B/privacy%7D