CVE-2026-61762

NVIDIA Megatron Bridge, an open-source framework for training and fine-tuning large language models, contains a deserialization-of-untrusted-data flaw (CWE-502) in which untrusted serialized input is deserialized without adequate validation. The attack vector is local (CVSS AV:L) with low privileges required, meaning an attacker must already be able to supply untrusted data to a Megatron Bridge process, such as a model checkpoint or other serialized artifact loaded during a training or restoration workflow. A successful exploit could allow arbitrary code execution, tampering with training data or results, and disclosure of sensitive information in the compute environment, with high confidentiality, integrity, and availability impact. Anyone running Megatron Bridge in training pipelines that load externally sourced or shared artifacts is potentially affected. As of this analysis there is no known public proof-of-concept, the flaw is not in CISA's KEV catalog, and EPSS puts 30-day exploitation probability at roughly 0.3% (18th percentile).

What to do: Check the NVIDIA security advisory for this CVE and upgrade Megatron Bridge to a fixed version as soon as it is published. Until then, avoid deserializing checkpoints or other model artifacts from untrusted or shared sources, and restrict who can place serialized files into training pipelines. Because there is no known exploitation, prioritize patching over emergency response, and review whether shared checkpoint stores (e.g., team or community model repos) feed into your runs.