Critical Path Traversal Vulnerability in Weights & Biases wandb: CVE-2026-91771

Executive Intelligence Brief

A critical path traversal vulnerability (CVE-2026-91771) has been identified in Weights & Biases wandb, a popular platform for machine learning and data science. This vulnerability affects versions before 0.29.0 and allows attackers to manipulate file names, potentially leading to code execution. With a CVSS score of 8.8, this vulnerability is considered high severity. Organizations using affected versions of wandb should immediately patch to version 0.29.0 or later to mitigate this risk.

Threat Overview

Weights & Biases wandb is a widely used platform in the machine learning and data science communities for tracking and visualizing model performance, data, and experiments. The platform's popularity stems from its ability to provide detailed insights and collaboration features, making it a critical tool for many organizations.

The vulnerability CVE-2026-91771 affects the File.download function in wandb versions before 0.29.0. This function fails to properly validate file names returned from server responses, allowing for path traversal attacks. An attacker controlling the backend can supply file names with directory traversal sequences (e.g., ../../) to write files outside the intended download directory.

This could potentially enable code execution through modification of shell startup files or Python import paths. The vulnerability has a significant impact due to wandb's widespread use in the machine learning and data science sectors.

Technical Deep Dive

Vulnerability Classification

The vulnerability is classified as CWE-22, which refers to Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'). This class of vulnerability occurs when a software application fails to properly sanitize or validate user-supplied input that is used to construct file paths.

The CVSS vector for this vulnerability is: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H. This breaks down into:
- AV:N: Attack Vector is Network
- AC:L: Attack Complexity is Low
- PR:N: Privileges Required is None
- UI:R: User Interaction is Required
- S:U: Scope is Unchanged
- C:H: Confidentiality Impact is High
- I:H: Integrity Impact is High
- A:H: Availability Impact is High

Root Cause Analysis

The root cause of this vulnerability lies in the File.download function of the wandb library. Specifically, the function fails to validate or sanitize file names obtained from server responses. This allows an attacker to provide file names that include directory traversal sequences, such as ../../, which can lead to writing files outside the intended directory.

Attack Vector & Chain

The attack vector involves an attacker controlling the backend being able to supply malicious file names. The attack chain includes:
1. The attacker prepares a malicious file name with directory traversal sequences.
2. The attacker sends the malicious file name to the wandb application.
3. The wandb application, without proper validation, uses the file name to construct a file path.
4. The attacker can then write files outside the intended download directory.

Exploitation Scenario Walkthrough

Scenario: Remote Code Execution via Path Traversal in wandb

Reconnaissance: An attacker discovers a target organization using a vulnerable version of wandb. This could be done through various means such as Shodan scans, exposed API documentation, or public GitHub repositories.

Weaponization: The attacker prepares a malicious file name with directory traversal sequences (e.g., ../../) that, when processed, could lead to writing a file in a location that allows code execution, such as a shell startup file or a Python import path.

Delivery & Exploitation: The attacker supplies the malicious file name to the wandb application's File.download function. The application, lacking proper validation, processes the file name and writes the file to a location outside the intended download directory.

Post-Exploitation: With the ability to write files to arbitrary locations, the attacker could modify shell startup files or add malicious Python modules to the import path. This could lead to code execution under the context of the wandb application or potentially escalate privileges.

Impact Realization: The final impact could be remote code execution, allowing the attacker to gain control over the system running the vulnerable wandb version. This could lead to further exploitation, such as data exfiltration, lateral movement within the network, or deployment of ransomware.

Exploitation in the Wild

There is no indication that this vulnerability is currently being actively exploited in the wild. However, given its high severity and the potential for remote code execution, it is likely that attackers will prioritize exploiting this vulnerability if they can identify vulnerable targets.

Impact Analysis

Direct Impact

The direct impact of this vulnerability is the potential for remote code execution. An attacker could write malicious files to arbitrary locations, leading to code execution under the context of the wandb application. This could result in confidentiality, integrity, and availability impacts.

Downstream & Cascading Effects

The downstream effects could include:
- Supply chain risk: If the vulnerable wandb library is used in a broader software supply chain, the impact could propagate to dependent systems and services.
- Regulatory implications: Depending on the nature of the data processed by wandb, there could be regulatory implications if confidentiality is breached.
- Operational disruption: Successful exploitation could lead to operational disruption if the attack results in denial of service or requires significant remediation efforts.

Affected Products & Versions

The vulnerability affects wandb versions before 0.29.0. The fixed version is 0.29.0 or later.

Detection & Threat Hunting

Indicators of Compromise

Indicators of compromise (IoCs) may include:
- Unusual file writes or modifications outside the expected wandb download directory.
- Unexpected changes to shell startup files or Python import paths.
- Anomalous network activity indicative of exploitation attempts.

Detection Rules & Signatures

Detection logic could involve monitoring for:
- Suspicious file system writes or modifications.
- Anomalies in network traffic that could indicate exploitation attempts.
- Changes to system files or user profiles that could indicate post-exploitation activity.

Threat Hunting Queries

Threat hunting queries could include searching for:
- Recent changes to files in sensitive directories.
- Unusual network connections to or from systems running wandb.
- Anomalies in system logs that could indicate exploitation or post-exploitation activity.

Remediation & Hardening

Immediate Actions (0-24 hours)

Organizations should immediately:
- Upgrade to wandb version 0.29.0 or later.
- Monitor for and restrict suspicious network traffic.
- Review system logs for signs of exploitation.

Short-Term Hardening (1-7 days)

In the short term, organizations should:
- Implement additional monitoring for file system changes.
- Consider network segmentation to limit the spread of potential exploitation.
- Review and harden configurations of systems running wandb.

Strategic Recommendations

Strategically, organizations should:
- Ensure robust vulnerability management practices to promptly apply patches.
- Implement comprehensive monitoring and logging to detect suspicious activity.
- Consider security controls such as web application firewalls (WAFs) to detect and prevent exploitation attempts.

Analyst Assessment

This vulnerability is considered high severity due to its potential for remote code execution and the widespread use of wandb in the machine learning and data science communities. While there is no indication of active exploitation, organizations should prioritize patching to version 0.29.0 or later. The likelihood of exploitation is expected to increase as attackers become aware of the vulnerability and its potential impact.

Sources

• National Vulnerability Database (NVD) - CVE-2026-91771