This report was translated by AI; nobody has reviewed the translation yet. Original language: Czech.
[TLP:CLEAR] NVIDIA fixes a high-severity vulnerability
NVIDIA fixes a high-severity vulnerability in its products [1]. Products and their fixed versions: NVIDIA Container Toolkit - 1.17.4 NVIDIA GPU Operator - 24.9.2
Due to a time-of-check time-of-use (TOCTOU) race condition in NVIDIA Container Toolkit and GPU Operator for Linux, an unauthenticated remote attacker can escalate privileges, potentially execute code or carry out a DoS attack [1][2]. The vulnerability bypasses the fix for the older CVE-2024-0132 [2][3] and is exploitable on systems using the default configuration, where a specially crafted container image can gain access to the host file system. Systems using CDI are not affected [1]. The fix for this vulnerability changes the default behavior of NVIDIA Container Toolkit: the CUDA Compatibility libraries (/usr/local/cuda/compat) are no longer automatically included in the container's default path, which may affect applications that depend on this behavior. To preserve it, the configuration option "allow-cuda-compat-libs-from-container" can be used, which enables this functionality again. However, enabling this option removes the protection against the vulnerability and is not recommended. Details on configuring this option can be found at [1]. For applications that depend on CUDA Forward Compatibility, the LD_LIBRARY_PATH variable can be modified to include /usr/local/cuda/compat. However, this approach may cause portability issues between driver versions [1].
The vulnerability is present in products:
- *:NVIDIA Container Toolkit;<=1.17.3
- *:NVIDIA GPU Operator;<=24.9.1
CVSS: CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:C/C:H/I:H/A:H
More information:
- CVE-2025-23359 at cve.org, NVD
- CWE-367: Time-of-check Time-of-use (TOCTOU) Race Condition at cwe.mitre.org
The vulnerability was publicly disclosed 11. 2. 2025.
Links
By CESNET, z.s.p.o. Michaela Ručková on 18. 2. 2025.