CVE-2026-27489: ONNX symlink path traversal

CISO Take

Any ML pipeline that loads ONNX model files from untrusted sources is exposed to arbitrary file read on the host system. A crafted model containing symlinks can leak credentials, secrets, or config files at load time — no code execution required from the attacker. Upgrade to ONNX 1.21.0 immediately; if you operate model registries or accept user-uploaded models, treat this as urgent.

What is the risk?

Risk is HIGH for organizations that load ONNX models from external sources (HuggingFace, model marketplaces, third-party vendors). The vulnerability is trivially exploitable — creating a malicious ONNX model with symlinks requires no specialized knowledge. The attack surface is broad: ONNX is the de facto interoperability format across PyTorch, TensorFlow, and virtually every production ML stack. Environments where model loading co-locates with cloud credentials, API keys, or internal config files face immediate credential exposure risk. CVSS is not yet scored, but the combination of low complexity, no auth required, and high-value targets in typical ML environments warrants HIGH priority.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
ONNX	pip	<= 1.20.0	`1.21.0`
21.1K OpenSSF 8.8 1.2K dependents Pushed 6d ago 91% patched ~44d to patch Full package profile →

Do you use ONNX? You're affected.

How severe is it?

CVSS 3.1

8.6 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 44% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AV Network

AC Low

PR None

UI None

S Changed

C High

I None

A None

What should I do?

6 steps

PATCH

Upgrade onnx pip package to >= 1.21.0 across all environments (dev, staging, prod). Run: pip install --upgrade onnx.
AUDIT

Identify all pipeline stages that call onnx.load() or equivalent — include CI/CD, model converters, serving layers.
SANDBOX

Load untrusted ONNX models in isolated environments (containers with read-only mounts, no secrets, minimal filesystem exposure).
VALIDATE PROVENANCE

Implement model signing or hash verification before loading; reject models from unverified sources.
DETECT

Alert on unexpected file access patterns from Python processes loading model files — specifically symlink resolution outside expected model directories.
SHORT-TERM WORKAROUND (if patching is blocked): Scan ONNX model archives for symlinks before loading using: python -c "import zipfile; z=zipfile.ZipFile('model.onnx'); [print(i) for i in z.infolist() if i.create_system==3]" (ONNX protobuf files embedded in zip containers).

What does CISA's SSVC say?

Decision Track*

Exploitation poc

Automatable Yes

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Supply Chain Data Extraction Framework Model AML.T0010.001 - AI Software AML.T0011.000 - Unsafe AI Artifacts AML.T0025 - Exfiltration via Cyber Means AML.T0037 - Data from Local System AML.T0058 - Publish Poisoned Models

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Art.9 - Risk management system

ISO 42001

6.1.2 - AI risk assessment 8.4 - AI system supply chain

NIST AI RMF

GOVERN-1.1 - Policies and procedures for AI risk management MAP-5.1 - AI supply chain risk identification

OWASP LLM Top 10

LLM03:2025 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2026-27489?

Any ML pipeline that loads ONNX model files from untrusted sources is exposed to arbitrary file read on the host system. A crafted model containing symlinks can leak credentials, secrets, or config files at load time — no code execution required from the attacker. Upgrade to ONNX 1.21.0 immediately; if you operate model registries or accept user-uploaded models, treat this as urgent.

Is CVE-2026-27489 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-27489, increasing the risk of exploitation.

How to fix CVE-2026-27489?

1. PATCH: Upgrade onnx pip package to >= 1.21.0 across all environments (dev, staging, prod). Run: pip install --upgrade onnx. 2. AUDIT: Identify all pipeline stages that call onnx.load() or equivalent — include CI/CD, model converters, serving layers. 3. SANDBOX: Load untrusted ONNX models in isolated environments (containers with read-only mounts, no secrets, minimal filesystem exposure). 4. VALIDATE PROVENANCE: Implement model signing or hash verification before loading; reject models from unverified sources. 5. DETECT: Alert on unexpected file access patterns from Python processes loading model files — specifically symlink resolution outside expected model directories. 6. SHORT-TERM WORKAROUND (if patching is blocked): Scan ONNX model archives for symlinks before loading using: python -c "import zipfile; z=zipfile.ZipFile('model.onnx'); [print(i) for i in z.infolist() if i.create_system==3]" (ONNX protobuf files embedded in zip containers).

What systems are affected by CVE-2026-27489?

This vulnerability affects the following AI/ML architecture patterns: model serving, training pipelines, MLOps/CI-CD model validation, model registries, on-device inference.

What is the CVSS score for CVE-2026-27489?

CVE-2026-27489 has a CVSS v3.1 base score of 8.6 (HIGH). The EPSS exploitation probability is 0.59%.

What is the AI security impact?

Affected AI Architectures

model servingtraining pipelinesMLOps/CI-CD model validationmodel registrieson-device inference

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0011.000 Unsafe AI Artifacts

AML.T0025 Exfiltration via Cyber Means

AML.T0037 Data from Local System

AML.T0058 Publish Poisoned Models

Compliance Controls Affected

EU AI Act: Art.9

ISO 42001: 6.1.2, 8.4

NIST AI RMF: GOVERN-1.1, MAP-5.1

OWASP LLM Top 10: LLM03:2025

What are the technical details?

Original Advisory

Open Neural Network Exchange (ONNX) is an open standard for machine learning interoperability. Prior to version 1.21.0, a path traversal vulnerability via symlink allows to read arbitrary files outside model or user-provided directory. This issue has been patched in version 1.21.0.

Exploitation Scenario

Attacker publishes a malicious ONNX model to a public model hub (HuggingFace, ONNX Model Zoo mirror, PyPI-hosted model package). Inside the model archive, they embed a symlink: model_weights -> /run/secrets/kubernetes.io/serviceaccount/token. A victim's automated ML pipeline — say a model validation job in a Kubernetes-based MLOps platform — downloads and calls onnx.load() on the model during validation. The ONNX library follows the symlink and reads the Kubernetes service account token, which the attacker can then use to authenticate against the cluster API. Alternatively, in a less cloud-native scenario: a data scientist downloads and loads the model locally, exposing ~/.aws/credentials or ~/.ssh/id_rsa. The attack requires zero interaction beyond distributing the model file — a realistic supply chain vector given how freely ONNX models are shared.

Weaknesses (CWE)

CWE-23 Relative Path Traversal Primary CWE-23 Relative Path Traversal Primary CWE-61 UNIX Symbolic Link (Symlink) Following Primary CWE-61 UNIX Symbolic Link (Symlink) Following Primary CWE-23 Relative Path Traversal CWE-61 UNIX Symbolic Link (Symlink) Following

CWE-23 — Relative Path Traversal: The product uses external input to construct a pathname that should be within a restricted directory, but it does not properly neutralize sequences such as ".." that can resolve to a location that is outside of that directory.

[Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
[Implementation] Inputs should be decoded and canonicalized to the application's current internal representation before being validated (CWE-180). Make sure that the application does not decode the same input twice (CWE-174). Such errors could be used to bypass allowlist validation schemes by introducing dangerous inputs after they have been checked. Use a built-in path canonicalization function (such as realpath() in C) that produces the canonical version of the pathname, which effectively removes ".." sequences and symbolic links (CWE-23, CWE-59). This includes: realpath() in C getCanonicalPath() in Java GetFullPath() in ASP.NET realpath() or abs_path() in Perl realpath() in PHP

Source: MITRE CWE corpus.