CVE-2026-105752: vLLM Package Vulnerability Leading to Cross-Tenant Cache Membership Oracle

vllm Updated 06 Oct 2026

Summary

06/10/2026
Yes i
Low
3.1

CVE-2026-105752: Affected - Ecosystem / package: pip / vllm - Affected versions: vLLM ≤ 0.25.1 (confirmed on 0.25.1, commit 752a3a504485). Affects vllm versions < 0.30.0. Fixed in 0.30.0. How to fix ↓

Free scan

Check if this vulnerability affects your application

Upload a manifest or SBOM and we'll show whether you're affected, plus every other risky package in the same scan.

Check my app free

No signup required · Results in about 60 seconds

Upload your dependency manifest or SBOM

Drag & drop your file here or

Java Javascript PHP Python Go Rust Ruby Dart C/C++ Elixir CycloneDX SPDX

Upload a supported dependency manifest or SBOM. No installation or source-code access required.

See supported files and SBOM formats
SBOM sbom.json · bom.json · bom.xml · *.spdx.json · *.cdx.json · *.cdx.xml
Java pom.xml · gradle.lockfile · buildscript-gradle.lockfile
Javascript package-lock.json · yarn.lock · pnpm-lock.yaml
PHP composer.lock
Python poetry.lock · requirements.txt · Pipfile.lock
Go go.mod
Rust Cargo.lock
Ruby Gemfile.lock
Dart pubspec.lock
C/C++ conan.lock
Elixir mix.lock
Enter your work email to receive your report link.

Verify your email to view the full report and start 30 days of continuous monitoring. No credit card required.

Sponsored Devnack Your last release was probably never penetration tested Manual testing by engineers who build software, not an auditor's checklist. Trusted by teams at Swift and Google Talk to Devnack

Background

vLLM is an open-source library designed for efficient and scalable language model serving, particularly in the context of GPT-OSS applications. It allows users to leverage built-in tools and multi-turn interactions for generating responses. However, the library has a vulnerability that affects its ability to isolate cached data between different tenants, potentially exposing sensitive information.

Vulnerability Detail

The vulnerability arises from the improper handling of the cache_salt parameter during tool continuations in the vLLM library. Specifically, when a multi-turn tool call is made, the initial request correctly includes the cache_salt, but subsequent tool continuations omit this parameter. This leads to the caching of sensitive data in a global unsalted namespace, allowing an attacker to infer whether a guessed prompt was processed by another tenant. The implications of this vulnerability are significant, as it undermines the isolation guarantees intended for multi-user environments, potentially exposing exact cached-token counts and enabling prompt membership oracle attacks.

How To Fix

Latest published version: 0.32.0 as of 22 Oct 2026

pip install --upgrade "vllm==0.32.0"

Upgrading to it may include breaking changes; the version below is the minimum fixed release for this vulnerability.

If you need the minimum change

0.30.0

This is the first release that contains the fix.

pip install --upgrade "vllm==0.30.0"

Tip — Not sure which version your app runs? Vulert can match this fix to it. Scan free →

Workaround

As a temporary workaround, avoid using the Harmony tool path for multi-turn interactions until the package is updated. If you must use it, ensure that sensitive data is not included in the prompts to minimize the risk of exposure.

Conclusion

The vLLM package has a critical vulnerability that affects its ability to maintain data isolation between tenants. By updating to the latest version and following the recommended fixes, users can mitigate the risks associated with this vulnerability. Regular monitoring and vulnerability assessments are essential for maintaining application security.

See every vulnerability recorded for vllm →

Share this vulnerability

FAQ

The vulnerability involves the dropping of the cache_salt during tool continuations, leading to potential cross-tenant data exposure.

Any authenticated tenant using the vLLM library in a shared deployment environment is at risk.

Attackers may be able to infer sensitive information about other tenants' prompts and cached data, undermining data isolation.

You can use Vulert's playground at Vulert Scanner to check for vulnerabilities in your application without needing to sign up.

The CVSS score for this vulnerability is 3.1, indicating a low severity level.

Related vulnerabilities

Recent vulnerabilities recorded for vllm.

How To Fix

Latest published version: 0.32.0 as of 22 Oct 2026

pip install --upgrade "vllm==0.32.0"

Upgrading to it may include breaking changes; the version below is the minimum fixed release for this vulnerability.

If you need the minimum change

0.30.0

This is the first release that contains the fix.

pip install --upgrade "vllm==0.30.0"

Tip — Not sure which version your app runs? Vulert can match this fix to it. Scan free →

Subscribe to Vulert on Telegram
Daily digest of new CVEs, critical and high severity, affected packages and fixes — plus a weekly roundup.
Join the channel →
Help improve this record
Report an issue or send feedback