CVE-2026-91771: Weights & Biases path traversal

CISO Take

Weights & Biases wandb before 0.29.0 does not validate file names returned by the server in File.download, so a malicious or compromised backend can write files outside the intended download directory. That write primitive can become code execution by overwriting shell startup files or Python import paths, which makes it a realistic route onto ML engineer workstations and training hosts that often hold cloud credentials and model artifacts. Exploitation likelihood is currently modest: EPSS is 0.01167 (top 34%), the CVE is not in CISA KEV, no public exploit or Nuclei template exists, and CISA SSVC says TRACK. The exploit condition matters, though: the attacker must control the backend or sit in a position to tamper with responses, and a user or pipeline must trigger a download. Self-hosted W&B servers and any tenant that lets untrusted users upload run files raise that exposure. Upgrade wandb to 0.29.0 or later, and until then avoid File.download against untrusted or self-hosted servers you do not fully control.

Sources:NVDEPSSATLASGitHub Advisoryvulncheck.com

What is the risk?

Medium-high. The CVSS of 8.8 reflects network reach, low complexity, no privileges and full CIA impact, but it requires user interaction and control of the backend or its responses. Real-world likelihood is lower than the score suggests: EPSS is 1.2%, there is no KEV listing, no public exploit and no scanner template. The highest exposure is in environments that connect wandb clients to self-hosted or third-party servers, or that run automated download jobs with elevated privileges. Package risk data (dependents, OpenSSF score) is unavailable, so blast radius is uncertain. Treat it as a supply-chain-style risk for ML tooling.

How does the attack unfold?

Initial Access

Attacker controls or compromises a W&B backend, or tampers with its responses, so file metadata returned to clients is malicious.

AML.T0010.001

User Execution

A data scientist or automated pipeline calls File.download on a file served by the malicious backend.

AML.T0011

Exploitation

The unvalidated file name with traversal sequences makes wandb write attacker content outside the download directory, such as into shell startup files or Python import paths.

AML.T0050

Impact

The payload runs on next shell or Python start, exposing credentials, datasets and model artifacts on the ML host.

Initial Access

Attacker controls or compromises a W&B backend, or tampers with its responses, so file metadata returned to clients is malicious.

AML.T0010.001

User Execution

A data scientist or automated pipeline calls File.download on a file served by the malicious backend.

AML.T0011

Exploitation

The unvalidated file name with traversal sequences makes wandb write attacker content outside the download directory, such as into shell startup files or Python import paths.

AML.T0050

Impact

The payload runs on next shell or Python start, exposing credentials, datasets and model artifacts on the ML host.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Weights & Biases	pip	—	No patch
11.3K1.3K dependentsPushed 4d ago0% patchedFull package profile →

Do you use Weights & Biases? You're affected.

How severe is it?

CVSS 3.1

8.8 / 10

EPSS

1.2%

chance of exploitation in 30 days

Higher than 67% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Moderate

What is the attack surface?

AVNetwork

ACLow

PRNone

UIRequired

SUnchanged

CHigh

IHigh

AHigh

What should I do?

1 step

1) Upgrade wandb to 0.29.0 or later (fix commit e52558f in the upstream repository) across developer machines, notebooks, containers, CI runners and training images; inventory with pip freeze or SBOM tooling. 2) Until patched, avoid File.download against self-hosted or untrusted W&B endpoints, and download into disposable sandboxes or containers with a read-only home directory and no write access to shell profiles or site-packages. 3) Run ML jobs as unprivileged users with least-privilege cloud credentials. 4) Restrict who can upload run files to shared W&B projects and verify the integrity of self-hosted W&B servers and their TLS endpoints. 5) Detect: monitor for file writes by python processes to shell startup files, site-packages or paths outside the target download directory, and for unexpected '../' sequences in file names logged by W&B API calls. 6) Rotate credentials on any host that downloaded files from an untrusted backend.

What does CISA's SSVC say?

DecisionTrack

Exploitationnone

AutomatableNo

Technical Impacttotal

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

How is it classified?

Code ExecutionSupply ChainFrameworkTraining DataAML.T0010.001- AI Software AML.T0011- User Execution AML.T0050- Command and Scripting Interpreter

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15-Accuracy, robustness and cybersecurity

ISO 42001

A.7.4-Quality of data for AI systems

NIST AI RMF

MANAGE 2.3-Responses to previously unknown risks and supply chain risks

OWASP LLM Top 10

LLM03-Supply Chain

Frequently Asked Questions

What is CVE-2026-91771?

Weights & Biases wandb before 0.29.0 does not validate file names returned by the server in File.download, so a malicious or compromised backend can write files outside the intended download directory. That write primitive can become code execution by overwriting shell startup files or Python import paths, which makes it a realistic route onto ML engineer workstations and training hosts that often hold cloud credentials and model artifacts. Exploitation likelihood is currently modest: EPSS is 0.01167 (top 34%), the CVE is not in CISA KEV, no public exploit or Nuclei template exists, and CISA SSVC says TRACK. The exploit condition matters, though: the attacker must control the backend or sit in a position to tamper with responses, and a user or pipeline must trigger a download. Self-hosted W&B servers and any tenant that lets untrusted users upload run files raise that exposure. Upgrade wandb to 0.29.0 or later, and until then avoid File.download against untrusted or self-hosted servers you do not fully control.

Is CVE-2026-91771 actively exploited?

No confirmed active exploitation of CVE-2026-91771 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-91771?

1) Upgrade wandb to 0.29.0 or later (fix commit e52558f in the upstream repository) across developer machines, notebooks, containers, CI runners and training images; inventory with pip freeze or SBOM tooling. 2) Until patched, avoid File.download against self-hosted or untrusted W&B endpoints, and download into disposable sandboxes or containers with a read-only home directory and no write access to shell profiles or site-packages. 3) Run ML jobs as unprivileged users with least-privilege cloud credentials. 4) Restrict who can upload run files to shared W&B projects and verify the integrity of self-hosted W&B servers and their TLS endpoints. 5) Detect: monitor for file writes by python processes to shell startup files, site-packages or paths outside the target download directory, and for unexpected '../' sequences in file names logged by W&B API calls. 6) Rotate credentials on any host that downloaded files from an untrusted backend.

What systems are affected by CVE-2026-91771?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, ML experiment tracking, model artifact registries, CI/CD for ML, notebook environments.

What is the CVSS score for CVE-2026-91771?

CVE-2026-91771 has a CVSS v3.1 base score of 8.8 (HIGH). The EPSS exploitation probability is 1.17%.

What is the AI security impact?

Affected AI Architectures

training pipelinesML experiment trackingmodel artifact registriesCI/CD for MLnotebook environments

MITRE ATLAS Techniques

AML.T0010.001AI Software

AML.T0011User Execution

AML.T0050Command and Scripting Interpreter

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.7.4

NIST AI RMF: MANAGE 2.3

OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

Weights & Biases wandb before 0.29.0 fails to validate the file name from server responses in the File.download function, allowing path traversal attacks. Attackers controlling the backend can supply file names with directory traversal sequences to write files outside the intended download directory, potentially enabling code execution through modification of shell startup files or Python import paths.

Exploitation Scenario

An attacker gains control of, or stands up a look-alike, W&B-compatible backend, for example a compromised self-hosted instance or a shared project that serves manipulated file metadata. A data scientist or CI job calls File.download on a run file. The malicious server returns a file name such as '../../.bashrc' or a path into site-packages. wandb writes the attacker's content outside the download directory. On the next shell start or Python import, the payload runs, giving the attacker access to the training host, cloud credentials, datasets and model artifacts.

Weaknesses (CWE)

CWE-22Improper Limitation of a Pathname to a Restricted DirectoryPrimary CWE-22Improper Limitation of a Pathname to a Restricted Directory

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory: 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.