Skip to main content

NVIDIA CUDA Toolkit: Nine Vulnerabilities in cuobjdump and nvdisasm — The Developer Tool Attack Surface

Bulletin: NVIDIA CUDA Toolkit — February 2025 | CVEs: CVE-2024-53870, CVE-2024-53872, CVE-2024-53873, CVE-2024-53874, CVE-2024-53875, CVE-2024-53878 (cuobjdump), CVE-2024-53877 and further nvdisasm CVEs | Disclosed: February 2025

Discovered by Palo Alto Networks Unit 42, via a dedicated month-long fuzzing campaign.

For executives

NVIDIA's CUDA Toolkit includes two utilities developers use to inspect compiled GPU code: cuobjdump and nvdisasm. In February 2025, Palo Alto Networks Unit 42 disclosed nine vulnerabilities across these two tools. An attacker who distributes a malicious compiled GPU binary — embedded in a shared model, a container image, or a development package — can trigger these vulnerabilities when a developer or automated pipeline runs either tool against the file. The attack surface is AI development pipelines, not production inference environments.

The most severe: CVE-2024-53873 — heap buffer overflow in cuobjdump

CVE-2024-53873 is an integer overflow during cubin file parsing. The parser calculates the size of a buffer to allocate, performs integer arithmetic that can overflow — wrapping to a small number — and the allocator creates a buffer that is too small. Subsequent writes produce a heap buffer overflow, potentially enabling code execution.

The broader cluster

The remaining cuobjdump CVEs are integer overflows and out-of-bounds read conditions in different cubin parsing code paths. The nvdisasm CVEs follow the same pattern: out-of-bounds reads triggered by malformed ELF input.

The attack scenario

Pre-trained models distributed via Hugging Face or similar repositories often include compiled CUDA kernels. Container images with baked-in GPU code are another vehicle. A CI/CD pipeline that automatically runs cuobjdump or nvdisasm as part of a model validation step is a particularly clean target.

Patch

Update to CUDA Toolkit version 12.5.1 on Windows or 11.9.2 on Linux.