Security researchers at JFrog have disclosed CVE-2026-105192, a critical vulnerability in LMCache, a KV-cache acceleration layer widely deployed alongside vLLM inference servers. The vulnerability carries a CVSS score of 9.8, making it one of the most severe AI infrastructure vulnerabilities disclosed this year.
The flaw affects all LMCache versions from 0.3.9 through the current stable release, 0.5.5. At the time of disclosure, no patched version was available, leaving administrators with self-hosted inference deployments with an urgent remediation challenge.
Technical Details
CVE-2026-105192 is an unauthenticated remote code execution vulnerability. Attackers can exploit the flaw without credentials, sending specially crafted requests to the LMCache service that allow arbitrary code execution on the host system.
LMCache serves as a KV-cache layer that accelerates LLM inference by caching computed key-value pairs across requests. The component has become a common addition to vLLM deployments, particularly in production environments serving high-throughput applications.
Immediate Mitigations
Until an official patch is released, administrators should implement the following mitigations:
First, ensure the LMCache service is not directly reachable from untrusted networks. The service should sit behind authentication at the network layer and should never be exposed to the public internet.
Second, implement network segmentation to isolate inference infrastructure from other systems. Even if LMCache is compromised, proper segmentation limits the blast radius of potential breaches.
Third, review existing deployments to identify whether LMCache is in use. Many organizations may have added the component as part of vLLM optimization guides without fully understanding the security implications.
Industry Impact
The vulnerability highlights the expanding attack surface of AI infrastructure. As organizations deploy more sophisticated inference stacks with multiple components, each additional dependency represents a potential entry point for attackers.
This is not an isolated incident. The AI security landscape has seen multiple vulnerabilities in inference infrastructure this year, including the Google ADK CVE-2026-79696 disclosed in September. Organizations running self-hosted inference should prioritize regular security assessments of their entire AI deployment stack.
The JFrog security team noted that the vulnerability was responsibly disclosed and that they are coordinating with the LMCache maintainers on a fix.