CVE-2026-61763

NVIDIA Megatron Bridge, a Python library in NVIDIA's NeMo ecosystem for scaling large-language-model training, contains a deserialization-of-untrusted-data flaw (CWE-502). An attacker with low-privileged local access could supply crafted serialized input — such as a malicious checkpoint or serialized object that the library loads — and have it processed by Megatron Bridge. A successful exploit could result in arbitrary code execution, tampering with training data or results, and disclosure of sensitive information on the affected host. Users and teams running Megatron Bridge in AI development or training environments are potentially affected, with risk concentrated on shared or multi-user systems where untrusted users can submit jobs or provide data files. No public proof-of-concept, KEV listing, or confirmed in-the-wild exploitation is known, and EPSS estimates only a 0.3% chance of exploitation in the next 30 days.

What to do: Check NVIDIA's security advisory (this CVE is CNA-assigned to NVIDIA PSIRT) for the affected and fixed version ranges and upgrade Megatron Bridge accordingly. Until patched, avoid deserializing checkpoints or other serialized artifacts from untrusted sources and restrict local access on hosts running the library, prioritizing shared multi-user systems and CI/training pipelines where untrusted users can submit jobs or data. Because the attack vector is local (CVSS AV:L/PR:L), internet-facing exposure is not the primary concern.