CVE-2026-115919

ONNX: Out-of-Bounds Read in RNN/GRU/LSTM Shape Inference (Old Opsets, Missing Rank Check)

Published Oct 26, 2026 CVSS 3.1 · 6.2 Medium EPSS < 1% KEV · Not listed
Description
Open Neural Network Exchange (ONNX) is an open standard for machine learning interoperability. Prior to 1.22.0, RNNShapeInference_opset1_to_6() processes malformed RNN and LSTM nodes from opsets 1 through 6 and GRU nodes from opsets 3 through 6 without verifying that the first input has rank three. The function accesses dim(0) and dim(1) on an undersized protobuf shape, which can read outside the intended repeated field and terminate an application when onnx.shape_inference.infer_shapes() processes an untrusted model that can pass onnx.checker.check_model(). This issue is fixed in version 1.22.0.

Analysis

AI-generated

Technical impact

The vulnerability is an out‑of‑bounds read in the shape inference routine for legacy RNN, LSTM, and GRU model nodes. When a malicious ONNX file with an undersized shape is processed, the function reads dimensions beyond the bounds of the repeated field, causing a crash of the application that invokes onnx.shape_inference.infer_shapes(). This leads to a denial of service without leaking sensitive data or compromising the integrity of the running process.

Affected Systems

The affected product is ONNX (onnx:onnx) and all releases before 1.22.0. RNN and LSTM nodes from opsets 1 through 6 and GRU nodes from opsets 3 through 6 are impacted.; users of ONNX 1.21.x and earlier must update to 1.22.0 or later.

Risk and Exploitability

The CVSS score of 6.2 indicates a moderate severity. The EPSS score is not available and the vulnerability is not listed in CISA’s KEV catalog. The attack vector is local or remote, depending on whether the target application loads untrusted ONNX models; an attacker can supply a crafted model file that, once checked by onnx.checker.check_model() and then sent to infer_shapes(), will trigger the crash. No direct remote code execution or privilege escalation is possible, but the application becomes unavailable to legitimate users.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Upgrade ONNX to version 1.22.0 or later
  • Validate or whitelist ONNX model files before loading them into the application
  • If an upgrade is not yet possible, remove or restrict calls to the out‑of‑bounds shape inference routine for legacy opsets

Generated by OpenCVE AI on October 26, 2026 at 18:22 UTC.

Tracking

Sign in to view the affected projects.

Advisories

Source ID Title
Github GHSA Github GHSA GHSA-48m4-24cq-37qf ONNX: Out-of-Bounds Read in RNN/GRU/LSTM Shape Inference (Old Opsets, Missing Rank Check)
History
Type Values Removed Values Added
Metrics ssvc

{'options': {'Automatable': 'no', 'Exploitation': 'none', 'Technical Impact': 'partial'}, 'version': '2.0.3'}

Type Values Removed Values Added
First Time appeared Onnx
Onnx onnx
Vendors & Products Onnx
Onnx onnx
Type Values Removed Values Added
Description Open Neural Network Exchange (ONNX) is an open standard for machine learning interoperability. Prior to 1.22.0, RNNShapeInference_opset1_to_6() processes malformed RNN and LSTM nodes from opsets 1 through 6 and GRU nodes from opsets 3 through 6 without verifying that the first input has rank three. The function accesses dim(0) and dim(1) on an undersized protobuf shape, which can read outside the intended repeated field and terminate an application when onnx.shape_inference.infer_shapes() processes an untrusted model that can pass onnx.checker.check_model(). This issue is fixed in version 1.22.0.
Title ONNX: Out-of-Bounds Read in RNN/GRU/LSTM Shape Inference (Old Opsets, Missing Rank Check)
Weaknesses CWE-125
References
Metrics cvssV3_1

{'score': 6.2, 'vector': 'CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H'}

Weaknesses

Sources

Status
PUBLISHED
Assigner
GitHub_M
Published
Updated
Oct 26, 2026, 7:34 PM
Reserved
Oct 23, 2026, 8:21 PM
Status
Awaiting Analysis
Published
Oct 26, 2026, 6:17 PM
Modified
Oct 26, 2026, 8:17 PM
Severity
Moderate
Public date
Oct 26, 2026, 5:24 PM