GHSA-q69g-f4f5-v8m7: Medium severity pip/onnx vulnerability
Summary
ONNX's Python tensor-materialization helpers could allocate memory according to attacker-controlled tensor dimensions without verifying that a packed sub-byte payload contained the declared number of elements.
A very small model could therefore cause a disproportionately large allocation when an application called onnx.numpyhelper.toarray() or constructed onnx.reference.ReferenceEvaluator with the model. The issue affected UINT2, INT2, UINT4, INT4, and FLOAT4E2M1 tensors stored in either rawdata or packed int32data.
Technical details
The vulnerable unpack2bit() and unpack4bit() paths first allocated a small array based on the available packed bytes, then resized it to the shape declared in TensorProto.dims. They did not first confirm that the payload encoded enough logical elements for that shape.
For example, the reporter demonstrated a 66-byte model containing one packed byte and a declared dimension of 700,000,000. The model passed onnx.checker.checkmodel() and allocated approximately 700 MB while materializing the initializer, followed by another large allocation when executing the reference evaluator.
Loading, checking, or running shape inference on the serialized model alone did not trigger the allocation. The vulnerable tensor had to be materialized.
Impact
An application that materializes packed tensors from an untrusted model can exhaust available memory, causing denial of service. Memory corruption and code execution were not demonstrated.
Affected versions
- Affected: ONNX through 1.21.0 - Patched: ONNX 1.22.0
Remediation
Upgrade to ONNX 1.22.0 or later.
The primary fix validates the logical element count before resizing the unpacked NumPy array. A complementary checker change rejects an undersized packed payload before materialization.
- Materialization fix: onnx/onnx#8033 - Checker defense-in-depth: onnx/onnx#8038
The five affected tensor types and both storage formats were re-tested against the published ONNX 1.23.2 wheel and are rejected before the oversized allocation.
To verify the behavior using a released vulnerable package in an isolated environment, install onnx==1.21.0 with python -m pip install "onnx==1.21.0". Run memory-exhaustion demonstrations only with an appropriate process memory limit.
Severity
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. Applications that expose ONNX model processing over a network should calculate an Environmental score with Modified Attack Vector set to Network.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
pip/onnxto a version that resolves this vulnerability.Fixed in 1.22.0 - Upgrade
Upgrade
ONNXto a version that resolves this vulnerability.Fixed in 1.22.0
Event History
Frequently Asked Questions
Which deployments are exposed in practice?
Applications are exposed when they process attacker-controlled ONNX models and either call onnx.numpy_helper.to_array() or create an onnx.reference.ReferenceEvaluator from the model. The affected tensor encodings are UINT2, INT2, UINT4, INT4, and FLOAT4E2M1 stored in raw_data or packed int32_data.
What does an attacker need to trigger the issue?
An attacker needs to provide a crafted model with packed sub-byte tensor data that is smaller than the number of elements declared by TensorProto.dims. Exploitation occurs when the application materializes the tensor or initializes the reference evaluator; the described example used a 66-byte model to induce an allocation of about 700 MB.
Can normal ONNX model validation identify malicious models?
Not reliably. The reported crafted model passed onnx.checker.check_model() despite declaring a dimension that the packed payload could not represent, so passing that check does not establish that sub-byte tensor payloads match their declared dimensions.
What is the likely operational impact?
The issue can cause disproportionately large memory allocations and denial of service. The affected paths can allocate once while materializing an initializer and again while executing the reference evaluator.