Severity by source
AV:L/AC:L/PR:N/UI:R/S:U/C:N/I:N/A:H
File-based local delivery with required user invocation of convert_version(); crash-only impact means C:N/I:N, PR:N because no credentials are needed to supply a file.
Primary rating from Vendor (https://github.com/onnx/onnx).
CVSS VectorVendor: https://github.com/onnx/onnx
Lifecycle Timeline
1Blast Radius
ecosystem impact- 6 pypi packages depend on onnx (6 direct, 0 indirect)
Ecosystem-wide dependent count for version 1.9.0.
DescriptionCVE.org
Summary
Null pointer dereference (SIGSEGV) in Upsample_6_7::adapt_upsample_6_7() (onnx/version_converter/adapters/upsample_6_7.h:31) when convert_version() processes a model with an Upsample node that has zero inputs. The adapter accesses node->inputs()[0]->sizes() without checking input count. 107-byte PoC crashes on Release build.
This is the same class of bug as the Cast adapter advisory (separate report) but in a different adapter, different file, and different operator.
Details
The Upsample 6→7 adapter validates attributes but not inputs:
// upsample_6_7.h:20-33
void adapt_upsample_6_7(..., Node* node) const {
ONNX_ASSERTM(
node->hasAttribute(width_scale_symbol) && node->hasAttribute(height_scale_symbol),
"...") // Attribute check PASSES
auto width_scale = node->f(width_scale_symbol);
auto height_scale = node->f(height_scale_symbol);
auto input_shape = node->inputs()[0]->sizes();
// ^^^^^^^^^^^^^^^^^^^^
// OOB when inputs().size() == 0 → SIGSEGV
}The PoC has an Upsample node at opset 6 with the required width_scale and height_scale attributes but zero inputs. The attribute assertions pass, then node->inputs()[0] on an empty ArrayRef:
- Release builds (
NDEBUG): bounds-check assertion compiled out → reads garbage pointer → SIGSEGV - Debug builds:
assert(Index < Length)atarray_ref.h:159→ SIGABRT
An Upsample node with zero inputs passes graphProtoToGraph() because the import code only resolves input names present in the protobuf.
PoC
import base64
import onnx
from onnx import version_converter
poc_b64 = "CAI6YQo8EgFZIghVcHNhbXBsZSoVCgt3aWR0aF9zY2FsZRUAAABAoAEBKhYKDGhlaWdodF9zY2FsZRUAAABAoAEBEgR0ZXN0YhsKAVkSFgoUCAESEAoCCAEKAggBCgIIBAoCCARCBAoAEAY="
model = onnx.load_from_string(base64.b64decode(poc_b64))
# CRASHES - Upsample_6_7 adapter dereferences empty inputs array
version_converter.convert_version(model, 7)
# SIGSEGV107-byte PoC. Confirmed SIGSEGV on both onnx 1.21.0 (pip) and 1.22.0 (source build).
Impact
Any application that uses onnx.version_converter.convert_version() on untrusted models is vulnerable. This includes model conversion pipelines and tools that auto-upgrade opset versions for compatibility. The crash is unrecoverable (SIGSEGV).
This vulnerability is part of a systemic pattern across multiple version converter adapters. A full audit of all ~45 adapters was performed as part of the fix; eight adapters were found with the same class of unguarded indexed access (cast_9_8, softmax_12_13, softmax_13_12, upsample_6_7, upsample_9_10, group_normalization_20_21, broadcast_forward_compatibility, upsample_9_8) and all have been fixed in PR #7813.
AnalysisAI
Null pointer dereference in the ONNX Python library's version converter crashes any application that calls onnx.version_converter.convert_version() on a crafted model containing an Upsample node with zero inputs, producing an unrecoverable SIGSEGV. Affected versions include at minimum onnx 1.21.0 and 1.22.0, confirmed by a 107-byte public proof-of-concept; the vulnerability is part of a systemic pattern spanning eight adapters. No public active exploitation is confirmed in CISA KEV, but the trivially small PoC eliminates any meaningful exploitation barrier for threat actors targeting ML model conversion pipelines.
Technical ContextAI
The ONNX Python package (pkg:pip/onnx) ships a C++ version-conversion framework with per-operator adapters that upgrade ONNX models between opset versions. The affected adapter Upsample_6_7::adapt_upsample_6_7() in onnx/version_converter/adapters/upsample_6_7.h handles the opset 6-to-7 transition for the Upsample operator. CWE-476 (Null Pointer Dereference) describes the root cause: the adapter validates the presence of width_scale and height_scale attributes but unconditionally dereferences node->inputs()[0] without first checking that the inputs array is non-empty. In release builds compiled with NDEBUG, LLVM ArrayRef's bounds assertion is elided, so an empty-array access reads a garbage pointer and faults with SIGSEGV; debug builds trigger SIGABRT via assert(Index < Length) at array_ref.h:159. The graph import path graphProtoToGraph() permits nodes with zero inputs because it only resolves input names explicitly listed in the protobuf, allowing the malformed node to pass graph loading and reach the adapter unchecked. The same unguarded indexed-access pattern was independently confirmed in seven additional adapters (cast_9_8, softmax_12_13, softmax_13_12, upsample_9_10, group_normalization_20_21, broadcast_forward_compatibility, upsample_9_8), all remediated together in PR #7813 and the associated fix PR #7916.
RemediationAI
Upgrade the onnx Python package to the first release that incorporates PR #7916 (https://github.com/onnx/onnx/pull/7916); check https://github.com/onnx/onnx/releases for the tagged release following that merge, as the exact patched version is not independently confirmed from available data - the v1.22.0 release tag is referenced but that version was confirmed vulnerable, so verify the changelog before treating it as fixed. As a compensating control where immediate upgrade is not feasible, restrict onnx.version_converter.convert_version() to models sourced exclusively from internal, trusted repositories; do not invoke the converter on user-uploaded, network-received, or third-party ONNX files. Running the conversion step in an isolated subprocess with a hard timeout and restart-on-crash logic provides crash containment at the cost of added latency and operational complexity, preventing a single malformed model from taking down a shared conversion service. Note that onnx.check_model() does not currently guard against this specific graph malformation and should not be relied upon as a standalone filter.
Wazuh SIEM platform versions 4.4.0 through 4.9.0 contain an unsafe deserialization vulnerability in the DistributedAPI t
BentoML version 1.4.2 and earlier contains an unauthenticated remote code execution vulnerability through insecure deser
pgAdmin 4 contains critical remote code execution vulnerabilities in the Query Tool download and Cloud Deployment endpoi
The renderLocalView function in render/views.py in graphite-web in Graphite 0.9.5 through 0.9.10 uses the pickle Python
BentoML is a Python library for building online serving systems optimized for AI apps and model inference. Rated critica
OpenSSL before 0.9.8za, 1.0.0 before 1.0.0m, and 1.0.1 before 1.0.1h does not properly restrict processing of ChangeCiph
pyLoad download manager version prior to 0.5.0b3.dev77 exposes the Flask SECRET_KEY through an unauthenticated endpoint.
Langflow (a visual LLM pipeline builder) contains a critical unauthenticated code execution vulnerability (CVE-2026-3301
In Mercurial before 4.1.3, "hg serve --stdio" allows remote authenticated users to launch the Python debugger, and conse
Unauthenticated remote code execution affects Kestra OSS (the open-source event-driven orchestration platform) prior to
Unauthenticated remote code execution in Marimo ≤0.20.4 allows attackers to execute arbitrary system commands via the `/
pyLoad is the free and open-source Download Manager written in pure Python. Rated medium severity (CVSS 5.3), this vulne
Same weakness CWE-476 – NULL Pointer Dereference
View allSame technique Denial Of Service
View allVendor StatusVendor
SUSE
Severity: ModerateShare
External POC / Exploit Code
Leaving vuln.today
EUVD-2026-42355
GHSA-hwpq-hmq9-wj77