AI Stack Alert Medium CVE-2026-115920 For AI Eng, AppSec, SOC / IR AI-assisted
ONNX <= 1.21.0: out-of-bounds read in ConvTranspose shape inference (CVE-2026-115920)
ONNX ConvTranspose shape inference reads out of bounds when the weight tensor has fewer than three dimensions and no kernel_shape attribute is set. A crafted model that passes onnx.checker.check_model() causes a SIGSEGV crash in onnx.shape_inference.infer_shapes(). The issue affects ONNX 1.2.1 through 1.21.0 and is fixed in 1.22.0.
Why it matters
Any application that runs shape inference on untrusted ONNX models can be crashed by a single crafted file, causing denial of service. The source reports no confidentiality or integrity impact.
Do this
- Upgrade ONNX to 1.22.0 or later in all environments that run shape inference.
- Identify services that call onnx.shape_inference.infer_shapes() on user-supplied models and prioritize them for upgrade.
- Until patched, reject or sandbox models from untrusted sources before shape inference, and run that step in an isolated worker process so a crash does not take down the main service.
- If the service exposes model processing over a network, score the issue with Modified Attack Vector set to Network, as the advisory recommends.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.18%
Signals
- NoCISA KEV
- 6.2CVSS
- 0.18%EPSS
- 28/100Priority
- 90%Confidence
Evidence
“A crafted model passed `onnx.checker.check_model()` and caused SIGSEGV when passed to `onnx.shape_inference.infer_shapes()`.”
“Affected: ONNX 1.2.1 through 1.21.0”
“Upgrade to ONNX 1.22.0 or later.”
Sources
Affected / related
ONNX, ConvTranspose shape inference out-of-bounds read
Published 2026-10-28 21:16 UTC · updated 2026-11-08 02:33 UTC · AI-assisted summary, checked against quoted sources · How we score