CVE-2026-61775

NVIDIA Megatron Bridge, NVIDIA's framework for training large language models, contains a flaw (CWE-502) in which it deserializes untrusted data without adequate validation. The attack vector is local: an attacker with low privileges on or access to data consumed by a system running Megatron Bridge would need to supply malicious serialized input, such as a crafted training artifact or checkpoint, that the library then deserializes. A successful exploit could give the attacker code execution on the training host, the ability to tamper with data, and access to sensitive information, which is reflected in the high-severity 7.8 CVSS score with high confidentiality, integrity, and availability impacts. Users affected are ML engineering and research teams running Megatron Bridge in development or training environments rather than internet-facing services. As of now there is no known public proof of concept, no CISA KEV listing, and the EPSS score of 0.2% indicates a low estimated probability of exploitation in the next 30 days.

What to do: Check NVIDIA's advisory (PSIRT) for the affected version range and upgrade Megatron Bridge to a fixed release as soon as it is published. In the meantime, restrict deserialization of untrusted checkpoints and other serialized inputs, verify the provenance of training artifacts, and limit low-privileged local access on hosts running Megatron Bridge training jobs.