NVIDIA AI Software Vulnerabilities and Patches March 2026

Affected Environment
NVIDIA B300 MCU, Megatron LM, Apex, Triton Inference Server, Model Optimizer, and NeMo Framework across supported platforms are impacted. Exposure depends on use of affected versions in your AI, GPU, or inference deployments and whether patches have been applied.

Threat Overview
Vulnerabilities include unsafe deserialization, RCE, input handling, and state corruption, leading to code execution, DoS, and data tampering. If unpatched, attackers could run code within NVIDIA components, affect system integrity, disrupt services, or access sensitive information.

Exposure Timeline
Issues were disclosed and documented on 25 March 2026; NVIDIA has released fixed versions or commits for affected products. There are currently no reports of active exploitation, but systems remain exposed until relevant updates are deployed.

Attack Surface
Exposed surfaces include MCU firmware, model loading, checkpoint and quantization workflows, HTTP endpoints, and deserialization paths. Risk is higher where NVIDIA AI tools are internet-facing, integrated into production pipelines, or handle untrusted files or inputs.

Technical Root Cause
Root causes include code injection opportunities, improper input validation, unsafe deserialization, and memory or state handling flaws. These weaknesses allow crafted files, payloads, or requests to influence execution flow or corrupt internal state in NVIDIA components.

Exploitation Pathway
Attackers may convince users to load malicious checkpoints, model files, or inputs in Megatron LM, NeMo, and Model Optimizer. For Triton and Apex, adversaries may send crafted HTTP or serialized payloads; B300 MCU issues arise from modifying unsupported registries.