GPU VulnDB

Database/AI/ML frameworks & serving

Hugging Face Accelerate: checkpoint parsing deserializes untrusted data into code execution

CVSS 7.8CVE-2025-14925AI/ML frameworks & servingcurated

Impact

Accelerate is the launcher and state-loader under a large share of multi-GPU training and distributed inference jobs, so this flaw sits on a path almost every fleet exercises. A crafted checkpoint loaded during resume or state restore executes attacker code in the worker process, with that worker's GPUs, mounts and credentials. On multi-node jobs the same poisoned state file is typically read by every rank, so one file compromises every worker in the job at once.

Who can reach it

No authentication against the library. The attacker needs write access to a checkpoint or state directory that a job resumes from, or the ability to get a job to load an untrusted checkpoint. Code runs as the job's user.

What to do

No fixed version is given in the record. Lock down write access to shared checkpoint directories - a resume path that any tenant can write to is the real exposure here - and prefer safetensors-based state. When a fixed release appears it is a pip upgrade and a restart of training launchers; no node drain or reboot.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.