CVE-2026-61766

NVIDIA Megatron Bridge (part of NVIDIA's NeMo/Megatron LLM training ecosystem) contains an insecure deserialization vulnerability (CWE-502) in which attacker-controlled serialized data is deserialized without adequate validation. Per the CVSS vector, exploitation requires local access with low privileges and no user interaction, consistent with scenarios such as loading a malicious checkpoint or training artifact into a Megatron Bridge job. A successful exploit could yield arbitrary code execution on the training host, tampering with data, and disclosure of sensitive information such as model weights or training datasets. Users are organizations and ML practitioners running Megatron Bridge for large-model training or fine-tuning; the affected and fixed version ranges are not stated in the available data, so the NVIDIA security advisory should be consulted. Exploitation status is quiet so far: no public proof-of-concept, not listed in CISA KEV, and EPSS puts 30-day exploitation probability at just 0.2% (12th percentile).

What to do: Inventory which training environments, containers and CI pipelines have Megatron Bridge installed, and compare those versions against NVIDIA's PSIRT advisory for CVE-2026-61766, upgrading to a patched release when published. As interim mitigation, avoid deserializing checkpoints, configs or other pickled artifacts from untrusted sources, and run training jobs with least-privilege accounts and filesystem/network restrictions to limit code-execution blast radius.