CVE-2026-61779

NVIDIA Megatron Bridge, a training framework in NVIDIA's NeMo ecosystem for large language models, contains a deserialization of untrusted data flaw (CWE-502), where serialized input is deserialized without adequate validation. The CVSS vector indicates the flaw is exploitable locally (AV:L) with low privileges and no user interaction, which in practice means a local user or process, or an attacker able to supply an untrusted checkpoint or other serialized artifact that the framework loads. A successful exploit could result in arbitrary code execution, tampering with data, or disclosure of sensitive information on the affected system. Anyone running Megatron Bridge to train or fine-tune models is exposed, with the highest risk in environments that load checkpoints or serialized state from external or untrusted sources. There is no known exploitation: no public proof of concept, the flaw is not in CISA's KEV, and EPSS estimates only a 0.2% probability of exploitation within 30 days.

What to do: Check NVIDIA's security advisory for this CVE to identify affected and patched Megatron Bridge versions and upgrade promptly. Until patched, avoid deserializing checkpoints, weights, or other serialized artifacts from untrusted or shared sources, and restrict which users can load data into training sessions. Because the attack vector is local, prioritize systems where untrusted users or multi-tenant workloads run Megatron Bridge.