Triton Inference Server DoS Proof-of-Concept (PoC)
This repository contains a malicious TensorRT engine file (model.plan) designed to demonstrate a Remote Denial of Service (DoS) vulnerability in the NVIDIA Triton Inference Server.
Vulnerability Overview
The vulnerability exists in the TensorRT backend of the Triton Inference Server. When the server attempts to deserialize a malformed engine file, it fails to properly validate the header metadata (specifically length-prefixed fields), leading to out-of-bounds memory access and a fatal process crash.
Files
model.plan: The malformed TensorRT engine file (Payload).
Steps to Reproduce
1. Requirements
Ensure you have the Triton client or TensorRT Python libraries installed:
