Executive Summary

This vulnerability is a segmentation fault in the OneFlow deep learning framework, specifically in the method oneflow._oneflow_internal.autograd.Function.FunctionCtx.mark_non_differentiable. When this method is called on a tensor outside the context of a custom autograd function, it causes unsafe memory access leading to a segmentation fault and program crash (core dump). Essentially, the function is misused outside its intended context, which triggers this fault. [1]

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can cause a Denial of Service (DoS) by crashing the program using OneFlow when the faulty function is called with crafted input. This means an attacker could cause the application to stop functioning by triggering the segmentation fault, leading to service interruptions. [1]

Useful 0 Not Useful 0

Detection Guidance

This vulnerability can be detected by running a test that calls the vulnerable function outside its intended context to see if it causes a segmentation fault. For example, executing the following Python code snippet in an environment with OneFlow v0.9.0 installed will reproduce the issue and cause a crash with a core dump: ```python import oneflow as flow ctx = flow._oneflow_internal.autograd.Function.FunctionCtx() ctx.mark_non_differentiable(flow.ones((1, 1))) ``` Monitoring for segmentation faults or core dumps when running this code indicates the presence of the vulnerability. [1]

Useful 0 Not Useful 0

Mitigation Strategies

Immediate mitigation involves avoiding calling `mark_non_differentiable` outside the context of a custom autograd function in OneFlow. Ensure that any use of `mark_non_differentiable` is strictly within a properly defined custom autograd function context to prevent unsafe memory access and segmentation faults. Additionally, consider updating OneFlow to a version where this issue is fixed once available. [1]

Useful 0 Not Useful 0