CVE-2026-61755

NVIDIA Megatron Bridge, NVIDIA's open-source framework for large language model training built on the Megatron/NeMo stack, contains a deserialization-of-untrusted-data flaw (CWE-502). An attacker with low-privileged local access (per the CVSS vector AV:L/PR:L) who can cause the framework to deserialize attacker-controlled or otherwise untrusted data can trigger the flaw without user interaction. A successful exploit may result in arbitrary code execution, tampering with data, and disclosure of sensitive information on the affected training environment. Anyone running NVIDIA Megatron Bridge in development, research, or production ML training environments is potentially affected, though the local attack vector means systems are not exposed in the way internet-facing services are. As of this analysis there is no known public proof-of-concept, it is not in the CISA KEV catalog, and EPSS estimates only a 0.3% chance of exploitation in the next 30 days.

What to do: Inventory environments where Megatron Bridge is installed (training clusters, CI pipelines, researcher workstations) and upgrade to the fixed version listed in NVIDIA's security bulletin, since specific versions are not provided in this data. As an interim mitigation, avoid deserializing checkpoints, artifacts, or serialized inputs from untrusted or third-party sources, and restrict local access on shared or multi-tenant training systems.