CVE-2026-61770

NVIDIA Megatron Bridge, NVIDIA's framework for large-scale LLM training in the NeMo/Megatron ecosystem, contains a deserialization-of-untrusted-data flaw (CWE-502). The CVSS vector (AV:L, PR:L) indicates exploitation requires local access or low privileges, consistent with an attacker supplying a malicious serialized artifact — such as a training checkpoint or other serialized input — that the framework deserializes. A successful exploit could allow arbitrary code execution, tampering with training data or artifacts, and disclosure of sensitive information on the training host or cluster. Users and operators running Megatron Bridge in AI training environments are affected, though no specific affected or fixed version ranges are provided in the available data. Exploitation status is currently quiet: there is no known in-the-wild exploitation, no public proof of concept, not in CISA KEV, and EPSS estimates only a 0.2% probability of exploitation in the next 30 days (11th percentile).

What to do: Upgrade Megatron Bridge to the fixed release identified in NVIDIA's security advisory, since no fixed version is specified in the available data. In the meantime, only deserialize artifacts (checkpoints, serialized state) from trusted sources, and restrict local/low-privileged access to environments that run Megatron Bridge training jobs.