Executive brief
NVIDIA TensorRT-LLM, a library used to optimize the performance of large language models, contains a flaw in its inference interface. An attacker can exploit this to consume all available GPU memory and processing power without restriction. This results in a denial-of-service, making the AI service unavailable to legitimate users and potentially disrupting broader business operations that rely on these models.
Technical details
A resource exhaustion vulnerability (CWE-770) exists in the OpenAI-compatible inference API of NVIDIA TensorRT-LLM. The flaw allows an attacker to trigger GPU resource allocations without proper limits or throttling mechanisms. While the attack vector is classified as local, it requires no specific privileges or user interaction to execute. A successful exploit enables an attacker to exhaust GPU memory or compute cycles, resulting in a complete denial of service (DoS) for the affected inference server. The vulnerability impacts versions up to and including v1.3.0 rc14.
Affected products
- NVIDIA TensorRT-LLM up to and including v1.3.0 rc14
Timeline
- 2026-07-14: advisory: Initial disclosure by NVIDIA