CVE-2026-130684
Description
A flaw was found in vLLM. When serving an untrusted machine learning model, a path traversal vulnerability allows a malicious model publisher to specify an arbitrary local file path in the model configuration. As a result, the service can be tricked into loading unintended local model checkpoints, allowing remote attackers to query private models and disclose sensitive model data through standard API requests.
Statement
This vulnerability is rated as Moderate impact because successful exploitation requires an administrator or automated pipeline to load an untrusted model repository, limiting the blast radius to environments serving third-party models alongside co-located private checkpoints. While vLLM parses the model configuration automatically during initialization without requiring remote code execution permissions, an attacker cannot directly force a running service to load arbitrary local paths without injecting a crafted configuration file. In typical isolated Red Hat OpenShift AI or containerized deployments, instances serving single models or isolated storage volumes significantly limit exposure to unauthorized local model disclosure.
Mitigation
Operational mitigation can be achieved by isolating storage volumes and inspecting model configurations prior to deployment:
- Isolate filesystem access: Run vLLM within dedicated containers or pods using restricted volume mounts, ensuring that only the specific model directory being served is accessible in the filesystem namespace. Restricting read access to co-located private checkpoints prevents the process from referencing unauthorized local model directories.
- Sanitize model configuration: Before launching vLLM with untrusted model artifacts, inspect the model's `config.json` and remove the `speculators_config` section or verify that `verifier.name_or_path` does not reference unexpected paths.
- Enforce network controls: Restrict access to vLLM API endpoints (such as `/detokenize` and inference endpoints) using network policies, ingress controllers, or firewall rules so that only trusted internal clients can query the service.
Caveats:
Removing the `speculators_config` entry disables speculative decoding capabilities for the affected model. Modifying volume mounts or configuration files requires restarting the vLLM container or service process.
Common Vulnerability Scoring System (CVSS) Score Details
Info alert:Important note
CVSS scores for open source components depend on vendor-specific factors (e.g. version or build chain). Therefore, Red Hat's score and impact rating can be different from NVD and other vendors. Red Hat remains the authoritative CVE Naming Authority (CNA) source for its products and services (see Red Hat classifications).
The following CVSS metrics and score provided are preliminary and subject to review.
CVSS v3 Score Breakdown
| Red Hat | NVD | cve.org | |
|---|---|---|---|
| Base Score | 5.4 | N/A | 5.4 |
| Attack Vector | Network | N/A | Network |
| Attack Complexity | Low | N/A | Low |
| Privileges Required | None | N/A | None |
| User Interaction | Required | N/A | Required |
| Scope | Unchanged | N/A | Unchanged |
| Confidentiality | Low | N/A | Low |
| Integrity Impact | Low | N/A | Low |
| Availability Impact | None | N/A | None |
Vector
Red Hat: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:L/I:L/A:N
cve.org: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:L/I:L/A:N
Understanding the Weakness (CWE)
Integrity,Confidentiality,Availability
Technical Impact: Execute Unauthorized Code or Commands
The attacker may be able to create or overwrite critical files that are used to execute code, such as programs or libraries.
Integrity
Technical Impact: Modify Files or Directories
The attacker may be able to overwrite or create critical files, such as programs, libraries, or important data. If the targeted file is used for a security mechanism, then the attacker may be able to bypass that mechanism. For example, appending a new account at the end of a password file may allow an attacker to bypass authentication.
Confidentiality
Technical Impact: Read Files or Directories
The attacker may be able to read the contents of unexpected files and expose sensitive data. If the targeted file is used for a security mechanism, then the attacker may be able to bypass that mechanism. For example, by reading a password file, the attacker could conduct brute force password guessing attacks in order to break into an account on the system.
Availability
Technical Impact: DoS: Crash, Exit, or Restart
The attacker may be able to overwrite, delete, or corrupt unexpected critical files such as programs, libraries, or important data. This may prevent the product from working at all and in the case of protection mechanisms such as authentication, it has the potential to lock out product users.
Frequently Asked Questions
Not sure what something means? Check out our Security Glossary.
Want to get errata notifications? Sign up here.