NVIDIA TensorRT: crafted TRT engine triggers excessive memory allocation and denial of service
Impact
A crafted TensorRT engine file makes TensorRT allocate an excessive amount of memory, which crashes or stalls the process that loads it. On a shared inference node, a service that accepts user-supplied engines can be knocked over, taking co-located models with it. The impact is availability only.
Who can reach it
Local, no privileges, but a user or service has to load an attacker-supplied TRT engine file.
What to do
Upgrade TensorRT to the fixed release named in NVIDIA bulletin 5921, then rebuild or redeploy the applications and containers that bundle it. The record does not give a fixed version. Until you upgrade, load engine files only from trusted sources.
References
Related entries
- NVIDIA License System - DLS virtual appliance: Installation scripts on the DLS appliance leave other users' credentialsCVE-2022-21818 · NVIDIA License System - DLS virtual applianceMedium
- Triton Inference Server: Insufficient access-control granularityCVE-2024-0103 · Triton Inference ServerMedium
- Intel Gaudi / gaudi-container-runtime: A path-traversal bug in the container runtime shim that wires Gaudi devices intoCVE-2026-32677 · Intel Gaudi / gaudi-container-runtimeMedium
- NVIDIA Windows GPU Display Driver (nvlddmkm.sys): Improper access control in the escape handler leaks informationCVE-2021-1055 · NVIDIA Windows GPU Display Driver (nvlddmkm.sys)Medium
- NVIDIA GPU Display Driver - Linux kernel module (nvidia.ko): An off-by-one error in nvidia.ko permits data tamperingCVE-2022-34684 · NVIDIA GPU Display Driver - Linux kernel module (nvidia.ko)Medium
- NVIDIA GPU Display Driver - Linux kernel module (nvidia.ko): An out-of-bounds array access in nvidia.ko givesCVE-2022-42254 · NVIDIA GPU Display Driver - Linux kernel module (nvidia.ko)Medium
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.