An attacker who can influence the code generation phase of NVIDIA TensorRT-LLM on Linux may inject or modify executable code because the library does not properly control the instructions it constructs. This vulnerability is a CWE‑829 flaw due to the improper restriction of operations in the generated code. If the attacker succeeds, the vulnerability can lead to arbitrary code execution, data tampering, and disclosure of sensitive information in the running process. The likely attack vector is inferred to be a process that receives crafted inputs to the library, which could be local or remote depending on how the application exposes the TensorRT-LLM interface.

All installations of NVIDIA TensorRT-LLM for Linux are potentially affected. No specific product versions are listed in the advisory, so any release containing the vulnerable code may be exposed. Users should verify their deployment against official NVIDIA release notes to confirm whether their version is impacted.

The CVSS score of 6.3 signals moderate to high impact, indicating that while the vulnerability is not trivial, its exploitation could have serious consequences. The EPSS score of less than 1% indicates that, as of this analysis, exploitation is unlikely, and the issue is not listed in CISA's KEV catalog. Nevertheless, the possibility of code execution, combined with the inferred path of manipulating input to the library, means that environments running TensorRT-LLM, especially if the library interfaces with untrusted data.