CVE-2026-61761

NVIDIA Megatron Bridge, a library in NVIDIA's NeMo/Megatron ecosystem for large language model training and post-training, contains a deserialization of untrusted data vulnerability (CWE-502). According to the CVSS vector (AV:L/PR:L/UI:N), an attacker with local access and low privileges can trigger the flaw without user interaction by causing the software to deserialize untrusted data. A successful exploit could yield code execution, tampering with data, and disclosure of sensitive information, with the impact rated high across confidentiality, integrity, and availability. Anyone running Megatron Bridge in development or training environments is potentially affected; because the attack vector is local, it is not exposed like an internet-facing service, limiting realistic exposure to hosts where the library is used. There is currently no known public proof-of-concept, no confirmed exploitation in the wild, and a low predicted exploitation probability (EPSS 0.3%, percentile 18).

What to do: Upgrade Megatron Bridge to the fixed release named in NVIDIA's security bulletin (the specific version numbers are not included in the available data, so verify against NVIDIA PSIRT's advisory). Until patched, restrict local access to hosts running Megatron Bridge, avoid deserializing untrusted or third-party checkpoints and serialized data, and check whether your training pipelines load externally supplied serialized artifacts. As the vector is local, standard host privilege hygiene and limiting who can run jobs in training environments reduce risk.