CVE-2026-61760

NVIDIA Megatron Bridge, NVIDIA's open-source library for scaling large-language-model training within the NeMo ecosystem, contains an unsafe deserialization vulnerability (CWE-502) in which untrusted data is deserialized. Per the CVSS vector (AV:L/AC:L/PR:L/UI:N), exploitation requires local access with low privileges and no user interaction, meaning an attacker or malicious process already running in a training environment must supply untrusted serialized data to the library. A successful exploit could lead to arbitrary code execution, tampering with data, and disclosure of sensitive information. Any organization running Megatron Bridge in LLM training or fine-tuning workflows is potentially affected, although the specific affected version ranges were not included in the available data and should be confirmed against NVIDIA's advisory. There is currently no evidence of exploitation: no public proof-of-concept exists, the flaw is not in CISA KEV, and EPSS estimates only a ~0.3% probability of exploitation within 30 days.

What to do: Check NVIDIA's PSIRT security bulletin for the affected version range and upgrade Megatron Bridge to the fixed release it specifies, since fixed versions are not stated in the available data. Until patching, avoid deserializing untrusted serialized inputs (e.g., third-party checkpoints or job configs) and restrict which users and processes can execute code on hosts running Megatron Bridge. Given the code-execution impact, monitor NVIDIA and community channels for the publication of a proof-of-concept.