This vulnerability arises from the use of unsafe serialized handles within NVIDIA TensorRT-LLM. The deserialization flaw, identified as CWE‑502, allows a malicious payload to be processed during model loading or inference, potentially resulting in arbitrary code execution, data tampering, and the disclosure of sensitive information. The affected library is available on any platform that supports TensorRT-LLM, meaning the flaw is not limited to a single operating system or architecture.

The flaw affects NVIDIA TensorRT‑LLM environments. No specific product version information is listed in this advisory, but the vulnerability applies to any installation of TensorRT‑LLM that processes serialized data.

The CVSS score of 6.3 indicates a moderate severity, and the EPSS score is not provided, making precise exploitation likelihood uncertain. The vulnerability is not currently flagged in CISA’s KEV catalog. Based on the nature of unsafe deserialization, it is likely that an attacker who can supply or influence serialized input—whether through a user‑controlled file, network message, or remote API—can induce the library to execute arbitrary code. The attack may be local if the library runs with elevated privileges or remote if it is exposed by an application that accepts untrusted data.