Executive brief
NVIDIA TensorRT-LLM, a library used to optimize the performance of large language models on Linux systems, contains a security flaw in its cluster management component. An attacker with local access to the system could manipulate the internal state of the computing cluster by sending unauthorized requests to an internal server. This could lead to the theft of sensitive information, unauthorized modification of data, or a complete disruption of the AI service.
Technical details
A vulnerability exists in the disaggregated orchestrator component of NVIDIA TensorRT-LLM for Linux due to missing authentication for critical functions (CWE-306). The flaw resides in the FastAPI server used by the orchestrator, which fails to properly validate or authenticate requests. A local attacker can exploit this by sending crafted HTTP requests to the FastAPI server to read, write, or delete internal cluster state information. Successful exploitation allows for unauthorized information disclosure, data tampering, and denial of service within the LLM infrastructure. The vulnerability affects versions up to and including v1.3.0 rc16.
Affected products
- NVIDIA TensorRT-LLM for Linux up to and including v1.3.0 rc16
Timeline
- 2026-07-14: advisory: NVIDIA published the vulnerability details.
- 2026-07-14: disclosed: CVE record published to the NVD dataset.