AI Stack Alert Medium CVE-2026-115922 For AI Eng, AppSec, SOC / IR AI-assisted
ONNX <= 1.21.0: DoS via unchecked expansion of packed sub-byte tensors (CVE-2026-115922)
ONNX versions through 1.21.0 allocate memory based on declared tensor dimensions when materializing packed sub-byte tensors (UINT2, INT2, UINT4, INT4, FLOAT4E2M1) without checking the payload size. A very small model can trigger a large allocation when passed to onnx.numpy_helper.to_array() or ReferenceEvaluator. The issue is fixed in ONNX 1.22.0.
Why it matters
Applications that load untrusted ONNX models with these Python APIs can be driven into memory exhaustion and denial of service. Memory corruption and code execution were not demonstrated.
Do this
- Upgrade ONNX to 1.22.0 or later in any service or pipeline that materializes tensors from untrusted models.
- Until upgraded, run model materialization and ReferenceEvaluator calls on untrusted models under a process memory limit.
- Inventory environments with onnx <= 1.21.0 using pip show onnx or a software bill of materials scan.
- If network-facing model processing exists, score the issue with Modified Attack Vector set to Network in the environmental assessment.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.13%
Signals
- NoCISA KEV
- 6.2CVSS
- 0.13%EPSS
- 28/100Priority
- 90%Confidence
Evidence
“A very small model could therefore cause a disproportionately large allocation when an application called `onnx.numpy_helper.to_array()` or constructed `onnx.reference.ReferenceEvaluator` with the model.”
“The vulnerable ONNX component processes the malicious model through a local library API or file boundary and is not itself bound to a network stack.”
“The primary fix validates the logical element count before resizing the unpacked NumPy array.”
Sources
Affected / related
onnx, packed sub-byte tensor expansion denial of service
Published 2026-10-28 21:16 UTC · updated 2026-11-08 02:33 UTC · AI-assisted summary, checked against quoted sources · How we score