Executive brief
NVIDIA TensorRT-LLM is a library used to optimize and accelerate the performance of large language models. A vulnerability in its inference interface allows an attacker to crash the service, leading to a denial of service. This could disrupt AI-powered applications and services relying on this library for real-time processing.
Technical details
A reachable assertion vulnerability (CWE-617) exists in the OpenAI-compatible inference API of NVIDIA TensorRT-LLM. The flaw is located within the sampler thread and can be triggered by a local attacker. By providing specific input that violates internal logic assumptions, an attacker can cause the program to terminate unexpectedly. This results in a denial of service (DoS) for the affected inference service. The vulnerability affects versions up to and including v1.3.0 rc15.
Affected products
- NVIDIA TensorRT-LLM <= v1.3.0 rc15
Timeline
- 2026-07-14: disclosed: Initial publication of CVE-2026-47475