Database/AI/ML frameworks & serving
NVIDIA Megatron-LM: unsafe deserialization in dataset loading allows code execution
Impact
Megatron-LM deserializes untrusted data when it loads a dataset, so a crafted dataset file runs code as the training process. On a shared training cluster, datasets often come from shared storage or outside sources, so anyone who can write to them can reach every job that reads them, along with those jobs' GPUs, credentials and outputs. NVIDIA rates the impact as code execution, information disclosure and data tampering.
Who can reach it
Local, low privilege: anyone who can place or modify a dataset that a Megatron-LM job loads. No user interaction beyond the normal load.
What to do
Upgrade Megatron-LM to the fixed release named in NVIDIA bulletin 5926 and restart the training jobs that use it. The record does not give a fixed version. Until you upgrade, restrict who can write to dataset locations.
References
Related entries
- MLflow: pickle evaluation artifact deserialization runs arbitrary code when a model is loadedCVE-2026-118623 · MLflow model artifact loading (PickleEvaluationArtifact)High
- MLflow: loading a SHAP explainer model deserializes untrusted data and runs arbitrary codeCVE-2026-118627 · MLflow SHAP explainer loading (load_explainer)High
- TorchServe: predictable /tmp extraction path lets a local user swap a model handler and run code as the service accountCVE-2026-123974 · PyTorch TorchServe (model archive extraction directory in /tmp)High
- Keras (`TorchModuleWrapper`): Unsafe deserialization of attacker-controlled PyTorch pickle inside a Keras modelCVE-2026-12484 · Keras (`TorchModuleWrapper`)High
- llama.cpp: integer overflow in GGUF file parsing lets a malicious model file run codeCVE-2026-126974 · llama.cpp (GGUF model file parser)High
- llama.cpp: heap buffer overflow in GGUF file parsing lets a malicious model file run codeCVE-2026-126975 · llama.cpp (GGUF model file parser)High
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.