CVE-2026-61757

NVIDIA Megatron Bridge contains a deserialization of untrusted data flaw (CWE-502) that could allow an attacker to achieve code execution, data tampering, or information disclosure. The flaw is triggered when the library deserializes attacker-controlled or untrusted serialized data — for example, loading untrusted checkpoints or similar artifacts — and the CVSS vector indicates a local, low-privilege attack context with no user interaction required. A successful exploit has high impact on confidentiality, integrity, and availability of the affected host. Anyone running NVIDIA Megatron Bridge, the NeMo-ecosystem library for post-training and fine-tuning large language models, is potentially affected, though the source data does not specify affected version ranges. There is no public proof-of-concept, no KEV listing, and no known in-the-wild exploitation, with EPSS estimating only about a 0.3% probability of exploitation in the next 30 days.

What to do: Upgrade Megatron Bridge to the latest patched release per NVIDIA's security bulletin (the fixed version is not specified in the available data). Until patched, avoid deserializing checkpoints or other serialized artifacts from untrusted sources and restrict local access to hosts running the library. There is no known exploitation, but ML pipelines that routinely load third-party checkpoints match the typical trigger for this flaw class, so treat patching as a priority.