CVE-2026-34447: ONNX symlink traversal reads host

CISO Take

Any pipeline or system loading untrusted ONNX model files is at risk of having sensitive host files read by a crafted model with symlinked external data — including secrets, SSH keys, or credentials accessible to the loading process. The attack surface is broad: CI/CD model validation pipelines, inference servers, and data scientists pulling models from public registries are all exposed. Patch to ONNX >= 1.21.0 immediately, and audit any workflow that loads ONNX models from external or untrusted sources.

What is the risk?

Medium severity on paper, but contextually higher in AI/ML environments. The CVSS 'local' vector is misleading: in ML pipelines, 'loading a model file' is routine and often automated — the attacker just needs to get a malicious ONNX file into the pipeline, which is trivial via public model hubs. No privileges required and the confidentiality impact is HIGH (full arbitrary file read as the process owner). In production environments, the inference service process may have access to API keys, database credentials, or internal certificates, making this a serious lateral movement enabler despite the medium score.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
ONNX	pip	< 1.21.0	`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

5.5 / 10

EPSS

0.2%

chance of exploitation in 30 days

Higher than 16% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

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 Local

AC Low

PR None

UI Required

S Unchanged

C High

I None

A None

What should I do?

6 steps

PATCH

Upgrade onnx to >= 1.21.0 immediately. Run pip install --upgrade onnx and validate with pip show onnx.
DETECT

Scan existing ONNX model files in your artifact stores for symlinks — find <model_dir> -type l before loading any model.
SANDBOX

Run model loading in isolated containers with restricted filesystem mounts (no access to /etc, /home, secrets volumes).
SUPPLY CHAIN

Verify ONNX model file provenance via checksums/signatures before loading in any pipeline.
MONITOR

Alert on unexpected file access patterns from ML inference processes (e.g., reads to /etc/passwd, ~/.ssh, credential directories).
WORKAROUND (pre-patch): Pre-scan model external data files for symlinks before loading; reject any model that contains symlinked external data references.

What does CISA's SSVC say?

Decision Track*

Exploitation poc

Automatable No

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.T0002.001 - Models AML.T0010.001 - AI Software AML.T0011.000 - Unsafe AI Artifacts AML.T0025 - Exfiltration via Cyber Means AML.T0037 - Data from Local System

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity of high-risk AI systems

ISO 42001

A.5.23 - Information security for use of cloud services / third-party AI components

NIST AI RMF

GOVERN 6.1 - Policies and procedures for AI supply chain risk management MANAGE 2.2 - Mechanisms to sustain AI risk management are in place

OWASP LLM Top 10

LLM03 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2026-34447?

Any pipeline or system loading untrusted ONNX model files is at risk of having sensitive host files read by a crafted model with symlinked external data — including secrets, SSH keys, or credentials accessible to the loading process. The attack surface is broad: CI/CD model validation pipelines, inference servers, and data scientists pulling models from public registries are all exposed. Patch to ONNX >= 1.21.0 immediately, and audit any workflow that loads ONNX models from external or untrusted sources.

Is CVE-2026-34447 actively exploited?

No confirmed active exploitation of CVE-2026-34447 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-34447?

1. PATCH: Upgrade onnx to >= 1.21.0 immediately. Run `pip install --upgrade onnx` and validate with `pip show onnx`. 2. DETECT: Scan existing ONNX model files in your artifact stores for symlinks — `find <model_dir> -type l` before loading any model. 3. SANDBOX: Run model loading in isolated containers with restricted filesystem mounts (no access to /etc, /home, secrets volumes). 4. SUPPLY CHAIN: Verify ONNX model file provenance via checksums/signatures before loading in any pipeline. 5. MONITOR: Alert on unexpected file access patterns from ML inference processes (e.g., reads to /etc/passwd, ~/.ssh, credential directories). 6. WORKAROUND (pre-patch): Pre-scan model external data files for symlinks before loading; reject any model that contains symlinked external data references.

What systems are affected by CVE-2026-34447?

This vulnerability affects the following AI/ML architecture patterns: model serving, training pipelines, MLOps CI/CD pipelines, model registries, edge deployments.

What is the CVSS score for CVE-2026-34447?

CVE-2026-34447 has a CVSS v3.1 base score of 5.5 (MEDIUM). The EPSS exploitation probability is 0.25%.

What is the AI security impact?

Affected AI Architectures

model servingtraining pipelinesMLOps CI/CD pipelinesmodel registriesedge deployments

MITRE ATLAS Techniques

AML.T0002.001 Models

AML.T0010.001 AI Software

AML.T0011.000 Unsafe AI Artifacts

AML.T0025 Exfiltration via Cyber Means

AML.T0037 Data from Local System

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.5.23

NIST AI RMF: GOVERN 6.1, MANAGE 2.2

OWASP LLM Top 10: LLM03

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, there is a symlink traversal vulnerability in external data loading allows reading files outside the model directory. This issue has been patched in version 1.21.0.

Exploitation Scenario

An adversary publishes a seemingly legitimate ONNX model to a public hub (Hugging Face, ONNX Model Zoo mirror, or internal registry). The model contains a large external data file declared in the ONNX protobuf, but the actual file on disk is a symlink pointing to /etc/passwd, ~/.aws/credentials, or a Kubernetes service account token at /var/run/secrets/kubernetes.io/serviceaccount/token. A CI/CD pipeline pulling and validating the model, or a data scientist running `onnx.checker.check_model('model.onnx')`, triggers the external data load. The process reads the symlink target and — depending on the application — logs it, stores it, or returns it in an API response. Combined with a second-stage exfiltration channel, the attacker retrieves cloud credentials and pivots to the production environment.

Weaknesses (CWE)

CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal') Primary CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal') Primary CWE-61 UNIX Symbolic Link (Symlink) Following Primary CWE-61 UNIX Symbolic Link (Symlink) Following Primary

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted 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
[Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.

Source: MITRE CWE corpus.