Executive brief
NVIDIA TensorRT-LLM, a library used to optimize the performance of Large Language Models (LLMs), contains a security vulnerability in how it processes data. An attacker with access to the local network could provide a specially crafted file that causes the system to crash or potentially allows unauthorized access to sensitive information. This could lead to service disruptions or the theft of proprietary data handled by the AI model.
Technical details
A heap-based buffer overflow vulnerability (CWE-122) exists in NVIDIA TensorRT-LLM during the tensor deserialization process. The flaw is triggered when the library processes a maliciously crafted tensor, leading to memory corruption. The attack vector is classified as 'Adjacent,' meaning the attacker must be on the same local network or subnet as the target. While the attack complexity is high and requires no user interaction or privileges, a successful exploit allows for arbitrary code execution, data modification, or a complete denial of service. The vulnerability affects versions up to and including v1.3.0 rc16.
Affected products
- NVIDIA TensorRT-LLM up to and including v1.3.0 rc16
Timeline
- 2026-07-14: disclosed: Initial publication of CVE-2026-47471 by NVIDIA