A critical vulnerability (CVE-2026–105192) was discovered in LMCache, an open-source add-on for vLLM used to serve AI models, allowing attackers to execute arbitrary code as root on affected servers. The vulnerability, present in versions 0.3.9 through 0.5.5 and some release candidates, exploited Python's pickle deserialization when LMCache was configured in its multiprocess mode with an exposed network port. While the LMCache team has released a fix in version 0.5.6, the incident highlights the need for robust network security and hardening when deploying shared services like KV caches in production environments, particularly on Kubernetes. AI
IMPACT Highlights critical security considerations for deploying AI model serving infrastructure like vLLM and its components.
RANK_REASON Security vulnerability discovered in an open-source add-on for AI model serving infrastructure.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →