Vulnerability in vLLM Inference Engine Allows Tenant Isolation Bypass

CVE-2026-105752

What is CVE-2026-105752?

The vLLM inference engine is vulnerable to an insecure cache management issue where prefix caching does not preserve the cache_salt value during continuation submissions. This design flaw permits an authenticated tenant to exploit cached tokens and infer whether a low-entropy post-tool history has been previously processed. Such behavior undermines the intended isolation between tenants and can lead to sensitive data leakage. The problem has been addressed in version 0.30.0, which enforces better cache handling and ensures tenant data remains secure.

References