CVE-2025-33255 Nvidia CVE debrief

Who should care

Organizations running NVIDIA TRT-LLM, especially deployments that use or expose the MPI server component, should treat this as relevant. Security teams, platform owners, and operators responsible for NVIDIA AI inference or serving stacks should verify whether any installed TRT-LLM instances are earlier than version 1.2 and review their exposure and access controls.

Technical summary

The supplied records describe a CWE-502 unsafe deserialization issue in NVIDIA TRT-LLM’s MPI server. The NVD data lists affected CPE coverage for nvidia:tensorrt_llm with versions ending before 1.2, and the CVSS v3.1 vector is AV:L/AC:H/PR:H/UI:N/S:C/C:H/I:H/A:H. In practical terms, successful exploitation could have high confidentiality, integrity, and availability impact, but the vector also indicates local access, high attack complexity, and high privileges are required.

Defensive priority

High. The severity is rated HIGH and the expected impact is broad, including code execution and loss of confidentiality, integrity, and availability. Prioritize inventory, version verification, and vendor-guided remediation for any TRT-LLM deployment that includes the MPI server, even though the CVSS vector suggests the exploit path is constrained by local access and high privileges.

Recommended defensive actions

• Inventory all NVIDIA TRT-LLM deployments and confirm whether any instance is running a version earlier than 1.2.
• Apply NVIDIA’s vendor guidance from the official advisory and move affected systems to a fixed version or other vendor-recommended remediation.
• Limit access to the MPI server component and reduce exposure of any local, high-privilege execution paths used by the service.
• Review operational logs and security telemetry for unusual MPI server activity or unexpected deserialization-related errors.
• Treat affected hosts as high-risk until patched and verify backups and recovery procedures in case tampering or service disruption occurs.

Evidence notes

Evidence in the supplied corpus includes the NVD analyzed record for CVE-2025-33255, the primary weakness classification CWE-502, and a vulnerable CPE range for nvidia:tensorrt_llm ending before 1.2. The CVSS v3.1 vector is explicitly provided as AV:L/AC:H/PR:H/UI:N/S:C/C:H/I:H/A:H, and the official reference list includes the NVD detail page, the CVE.org record, and NVIDIA’s vendor advisory URL. The corpus does not include the full advisory text, so remediation details are limited to the official reference links and the version boundary present in the record.

Official resources