Executive brief
OneFlow is an open-source deep learning framework used for building and scaling artificial intelligence models. A flaw in how the software handles GPU device requests allows an attacker to crash the application by providing an invalid hardware ID. This results in a denial-of-service (DoS) condition, potentially disrupting AI training workloads or production inference services.
Technical details
A device-ID validation flaw exists in OneFlow v0.9.0 within the CUDA synchronization logic. The vulnerability is triggered when the `flow.cuda.synchronize()` function is called with an invalid or out-of-range GPU device index. Internally, the framework fails to validate the index before passing it to `cudaSetDevice()`. When the underlying CUDA call fails, the application triggers a `Check failure` via Google Logging (glog), leading to an immediate abort and core dump. An attacker capable of influencing the device index parameter can remotely or locally cause the deep learning process to terminate unexpectedly.
Affected products
- OneFlow-Inc OneFlow 0.9.0
Timeline
- 2025-06-30: disclosed: Issue reported on GitHub repository
- 2026-01-28: advisory: CVE published to NVD