Executive brief
PyTorch, a popular machine learning framework, is vulnerable to a local denial-of-service attack. An attacker with local access to a system running PyTorch can trigger a crash by providing specific parameters to a pooling function. This could lead to the interruption of machine learning workloads or service outages on affected systems.
Technical details
A vulnerability in PyTorch versions prior to 2.7.1-rc1 allows a local attacker to cause a denial of service (DoS) via a floating point exception. The issue resides in the 'torch.mkldnn_max_pool2d' function, specifically when handling certain input parameters such as a 'stride' value of zero. This lack of proper error checking in the MKLDNN (oneDNN) integration leads to a process crash. Exploitation requires the ability to execute local Python code using the affected library. A fix is available in version 2.7.1-rc1.
Affected products
- PyTorch PyTorch < 2.7.1-rc1
Timeline
- 2025-03-16: disclosed: Issue reported on GitHub
- 2025-03-30: advisory: GitHub and VulDB advisories published