GPU VulnDB

Database/AI/ML frameworks & serving

Hugging Face Transformers: megatron_gpt2 checkpoint parsing deserializes untrusted data into code execution

CVSS 7.8CVE-2025-14924AI/ML frameworks & servingcurated

Impact

Loading a crafted Megatron-GPT2 checkpoint runs attacker code in the loading process. On a GPU node that is the training or serving worker itself, with its GPU devices, mounted dataset and checkpoint shares, and any Kubernetes service account or cloud credential the pod carries. Fleets that ingest externally produced Megatron checkpoints - a common path when converting weights from another training stack - take this on every conversion run. ZDI reports version 4.55.0 as affected.

Who can reach it

No authentication needed against the library; the attacker must supply a checkpoint that a process loads. In practice: any tenant or pipeline that can place a file in a watched model directory or registry that conversion jobs read.

What to do

The record names no fixed version. Until one lands, run checkpoint conversion in an isolated container with no cluster credentials and no write access to shared model stores, and allowlist the sources jobs may load from. The eventual fix is a pip upgrade plus a restart of the affected training or inference processes.

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.