Out-of-bounds read in onnx - CVE-2026-115920

Vulnerability details

The vulnerability allows a remote attacker to cause a denial of service by crashing an application performing ONNX shape inference.

The vulnerability exists due to an out-of-bounds read in ConvTranspose shape inference when an undersized weight tensor produces an empty effective_kernel_shape vector that is subsequently indexed. A remote attacker can supply a crafted model with a weight tensor having fewer than three dimensions and no kernel_shape attribute to trigger a process crash when the model is processed through the local shape inference API.

Affected software

How to mitigate CVE-2026-115920

External References

• https://github.com/onnx/onnx/security/advisories/GHSA-6pwm-x34p-fvj2
• https://github.com/onnx/onnx/commit/2ec288908952eea9dcf48d4b197915e1f72c6abf

Related Security Bulletins

• Multiple vulnerabilities in onnx