AI Stack Alert Medium CVE-2026-115919 For AI Eng, AppSec AI-assisted
ONNX through 1.21.0: out-of-bounds read in RNN/GRU/LSTM shape inference (CVE-2026-115919)
ONNX shape inference reads tensor dimensions without a rank check for RNN, GRU, and LSTM nodes in opsets 1–6 when the first input rank is not three. A crafted model can terminate a process that calls infer_shapes(), and the model passes onnx.checker.check_model(). The issue is fixed in ONNX 1.22.0.
Why it matters
Any service that runs ONNX shape inference on untrusted models can be crashed by one malformed file, which is a denial-of-service risk. No confidentiality or integrity impact was demonstrated in the source.
Do this
- Upgrade the onnx package to 1.22.0 or later in all environments that process models.
- Inventory services that call onnx.shape_inference.infer_shapes() on user-supplied or third-party models.
- Until upgraded, reject or quarantine untrusted models that contain RNN, GRU, or LSTM nodes at opsets 1–6.
- Run untrusted model processing in isolated worker processes with automatic restart so a crash does not take down the wider service.
- After upgrading, confirm malformed-model handling raises InferenceError rather than crashing the process.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.18%
Signals
- NoCISA KEV
- 6.2CVSS
- 0.18%EPSS
- 27/100Priority
- 85%Confidence
Evidence
“A crafted model could therefore terminate a process that called `onnx.shape_inference.infer_shapes()`.”
“Upgrade to ONNX 1.22.0 or later. The fix validates that the first input has rank three before accessing its dimensions.”
“The vulnerable schemas are limited to RNN and LSTM opsets 1–6 and GRU opsets 3–6.”
Sources
Affected / related
ONNX, Out-of-bounds read in shape inference
Published 2026-10-28 21:16 UTC · updated 2026-11-08 02:34 UTC · AI-assisted summary, checked against quoted sources · How we score