# CVE-2026-115922: ONNX memory exhaustion in packed tensor unpacking

Source: https://junglewise.ai/threats/cve-2026-115922-open-neural-network-exchange-onnx-is-an-open-standard-for  
Publisher: Junglewise Threat Intelligence, https://junglewise.ai/threats  
Published: 2026-10-26. Updated: 2026-10-26  
License: CC BY 4.0 with a link back

CVE-2026-115922 · Severity: medium · CVSS 6.2 · Published 2026-10-26

Open Neural Network Exchange (ONNX) is a standard for sharing machine learning models across different frameworks. When loading an untrusted ONNX model that contains specially crafted packed tensors (a compressed data format), an application can be forced to allocate excessive memory, causing the application to crash or become unresponsive. This denial-of-service attack occurs when processing tensor models with mismatched payload sizes.

Technologies: [ONNX](https://junglewise.ai/threats/technologies/onnx-onnx), [onnx (PyPI)](https://junglewise.ai/threats/technologies/onnx). Vendors: [ONNX](https://junglewise.ai/threats/vendors/onnx), [PyPI](https://junglewise.ai/threats/vendors/pypi).

## Quick answers

**Is CVE-2026-115922 exploited in the wild?** No. Junglewise Threat Intelligence has no report of CVE-2026-115922 being exploited in the wild.

**How severe is CVE-2026-115922?** Medium severity, with a CVSS score of 6.2, published 26 October 2026.

**Which products and versions are affected?** ONNX before 1.22.0; onnx (PyPI) <= 1.21.0, ECOSYSTEM:introduced=0,fixed=1.22.0.

**Is there a fix for CVE-2026-115922?** Yes. onnx (PyPI): upgrade to 1.22.0 or later. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.

**How is CVE-2026-115922 attacked?** It is attacked over the network, in a software library.

**Am I affected by CVE-2026-115922?** Check whether you run ONNX or onnx (PyPI) at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

## How to fix

- onnx (PyPI): upgrade to 1.22.0 or later

  ```sh
  pip install "onnx>=1.22.0"
  ```


## Technical details

The _unpack_2bit() and _unpack_4bit() functions in ONNX fail to validate that raw_data or int32_data fields contain sufficient bytes to match the declared tensor dimensions. An attacker can craft a TensorProto with small payload but large declared element count, causing the numpy_helper.to_array() or ReferenceEvaluator materialisation code to allocate disproportionately large memory. The vulnerability affects UINT2, INT2, UINT4, INT4, and FLOAT4E2M1 tensor types.

## Affected products

- ONNX ONNX before 1.22.0

## Identifiers

- CVE-2026-115922
- GHSA-q69g-f4f5-v8m7

## Timeline

- 2026-10-26: disclosed

## References

- [https://github.com/onnx/onnx/commit/05c1bc3bceceb8d3cf11e9e7fc7f19f971c44988](https://github.com/onnx/onnx/commit/05c1bc3bceceb8d3cf11e9e7fc7f19f971c44988)
- [https://github.com/onnx/onnx/commit/8de2cf6aa47928e6515ac6d9e1a8b4758886a7df](https://github.com/onnx/onnx/commit/8de2cf6aa47928e6515ac6d9e1a8b4758886a7df)

## Related threats

- [ONNX null pointer dereference in shape inference](https://junglewise.ai/threats/cve-2026-115921-open-neural-network-exchange-onnx-is-an-open-standard-for)
- [ONNX ConvTranspose shape inference denial of service](https://junglewise.ai/threats/cve-2026-115920-open-neural-network-exchange-onnx-is-an-open-standard-for)
- [ONNX out-of-bounds read in RNN shape inference](https://junglewise.ai/threats/cve-2026-115919-open-neural-network-exchange-onnx-is-an-open-standard-for)
- [ONNX TOCTOU race condition in save_external_data](https://junglewise.ai/threats/onnx-toctou-race-condition-in-save-external-data-b85422b8)
- [ONNX save_external_data symlink-following file write](https://junglewise.ai/threats/cve-2026-49114-onnx-save-external-data-symlink-following-file-write)