CVE-2026-65119

Untrusted Tokenizer Loading in TensorRT-LLM May Allow Code Execution

Published Oct 27, 2026 CVSS 3.1 · 8.4 High EPSS < 1% KEV · Not listed
Description
NVIDIA TensorRT-LLM for any platform contains a vulnerability in KvCacheAwareRouter tokenizer loading, where an attacker could cause inclusion of functionality from an untrusted control sphere. A successful exploit might lead to code execution, information disclosure, data tampering, and denial of service.

Analysis

AI-generated

Technical impact

The flaw resides in the KvCacheAwareRouter tokenizer loader of NVIDIA TensorRT‑LLM. This vulnerability is a CWE‑829 weakness, meaning it allows loading executable code from untrusted paths. An attacker that can influence the tokenizer loading path could force the execution of code originating from an untrusted control sphere. Successful exploitation can lead to arbitrary code execution, unauthorized information disclosure, data tampering, or denial of service within the affected TensorRT‑LLM environment.

Affected Systems

Affected vendors and products are NVIDIA TensorRT‑LLM running on any supported platform. The CVE does not list specific software version numbers, so all releases of TensorRT‑LLM provided by NVIDIA before the public patch are considered vulnerable until the control of tokenizer loading is hardened.

Risk and Exploitability

The vulnerability carries a CVSS score of 8.4, indicating high severity and a substantial impact should exploitation succeed. EPSS information is not available, and the issue is not yet in the CISA KEV catalog, but the high score suggests a non‑trivial risk. The attack surface is inferred to involve the ability to supply or modify tokenizer configuration files; an attacker with local or remote access to the system could potentially trigger the vulnerability by placing malicious tokenizer modules in an untrusted location, allowing the loader to execute them. Monitoring for abnormal loader activity and restricting access to tokenizer resources is recommended.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Upgrade to the latest NVIDIA TensorRT‑LLM release that incorporates the tokenizer loading fix.
  • Restrict tokenizer module loading to directories containing only signed, trusted files; block untrusted control spheres.
  • Validate tokenizer files by checking integrity hashes or signatures before loading them into the KvCacheAwareRouter.

Generated by OpenCVE AI on October 27, 2026 at 14:40 UTC.

Tracking

Sign in to view the affected projects.

Advisories

No advisories available
History
Type Values Removed Values Added
Metrics ssvc

{'options': {'Automatable': 'no', 'Exploitation': 'none', 'Technical Impact': 'total'}, 'version': '2.0.3'}

Type Values Removed Values Added
First Time appeared Nvidia
Nvidia tensorrt-llm
Vendors & Products Nvidia
Nvidia tensorrt-llm
Type Values Removed Values Added
Title Untrusted Tokenizer Loading in TensorRT-LLM May Allow Code Execution
Type Values Removed Values Added
Description NVIDIA TensorRT-LLM for any platform contains a vulnerability in KvCacheAwareRouter tokenizer loading, where an attacker could cause inclusion of functionality from an untrusted control sphere. A successful exploit might lead to code execution, information disclosure, data tampering, and denial of service.
Weaknesses CWE-829
References
Metrics cvssV3_1

{'score': 8.4, 'vector': 'CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H'}

Weaknesses
  • CWE-829

    Inclusion of Functionality from Untrusted Control Sphere

Sources

Status
PUBLISHED
Assigner
nvidia
Published
Updated
Nov 4, 2026, 4:53 PM
Reserved
Jul 21, 2026, 5:12 PM
Status
Awaiting Analysis
Published
Oct 27, 2026, 1:16 PM
Modified
Nov 4, 2026, 5:17 PM