CVE-2026-61773

NVIDIA Megatron Bridge, NVIDIA's open-source library for large language model training and post-training, contains an insecure deserialization flaw (CWE-502) in which untrusted serialized data is deserialized without adequate validation. The vulnerability is triggered when the framework processes attacker-controlled serialized input, most likely a malicious checkpoint or other serialized artifact loaded into a training pipeline; the CVSS local attack vector with low privileges indicates the attacker also needs some foothold in the environment. A successful exploit could yield arbitrary code execution, tampering with training data or artifacts, and disclosure of sensitive information such as datasets, model weights, or credentials in the training environment. Users affected are teams running NVIDIA Megatron Bridge for LLM training or fine-tuning, particularly in shared or multi-tenant clusters where checkpoints or data files from outside sources are ingested. There is currently no known public proof-of-concept, the flaw is not in the CISA KEV catalog, and EPSS assigns only a 0.2% probability of exploitation within 30 days.

What to do: Check NVIDIA's security advisory (issued by NVIDIA PSIRT) for the affected version range and upgrade Megatron Bridge to the patched release it specifies. Until patched, avoid deserializing checkpoints or serialized artifacts from untrusted sources and limit low-privileged local access to training hosts. Because the attack vector is local, prioritize environments where untrusted users, jobs, or third-party checkpoints can feed serialized data into training runs.