Executive brief
NVIDIA NeMo is a framework for building conversational AI and speech models. A maliciously crafted model_config.yaml file can inject unsafe parameters during dataset loading, allowing an attacker to execute arbitrary code, tamper with data, cause service outages, or steal sensitive information. This risk applies when users load untrusted model configuration files.
Technical details
The vulnerability exists in NeMo's dataset-loading workflow where user-supplied model_config.yaml files are processed without adequate validation, allowing unsafe parameter injection. An attacker can craft a malicious configuration file to achieve code execution, data modification, denial of service, or information disclosure. The attack requires an attacker to provide a crafted model configuration that a user or system loads, but does not require prior authentication.
Affected products
- NVIDIA NeMo
Timeline
- 2026-09-22: disclosed