Critical Unpatched LMCache RCE Threatens LLM Inference Infrastructure
A critical, unpatched vulnerability in LMCache allows unauthenticated attackers to execute remote code on the LLM inference cache layer. The flaw, tracked as CVE-2026-105192, carries a CVSS score of 9.8 and remains unpatched as of this writing.
What Happened: JFrog Discloses Unpatched LMCache Zero-Day Under Active Threat
On October 7, 2026, JFrog’s security research team disclosed a critical remote-code-execution vulnerability in LMCache, an open-source key-value cache layer used to accelerate large language model (LLM) inference in engines such as vLLM. The flaw affects every version from 0.3.9 through 0.5.5, plus current 0.5.6 release candidates and the development branch.
LMCache’s maintainers have not yet published a security advisory or a fixed release. Consequently, any deployment running the multiprocess server on a routable address is at immediate risk of full system compromise.
Technical Details of the CVE-2026-105192 Vulnerability
The vulnerability resides in LMCache’s multiprocess mode, also called distributed mode. In this configuration, LMCache opens an unauthenticated ZeroMQ ROUTER socket on port 5555 so worker processes can register and share KV cache blocks. The socket has no CURVE, ZAP, password, or message authentication.
Messages on the socket are encoded with msgpack. Extension code 1 is registered for DeviceIPCWrapper and is passed to DeviceIPCWrapper.Deserialize, which in turn calls Python’s pickle.loads. Because the deserialization occurs while the server is still decoding REGISTER_KV_CACHE arguments, before any handler logic runs, a single crafted ZeroMQ DEALER message can execute arbitrary code as the LMCache user. On the project’s official container images, that user is root.
Affected Versions and Deployment Types
- Affected versions: 0.3.9 through 0.5.5, plus 0.5.6 release candidates and the dev branch
- Default single-process mode: Not vulnerable (no ZeroMQ socket opened)
- Multiprocess mode with localhost bind: Vulnerable locally, but not remotely reachable
- Multiprocess mode with routable address (e.g.,
--host 0.0.0.0): Remotely exploitable without authentication
Exploitation Requirements
- LMCache running in multiprocess mode
- ZeroMQ transport bound to a routable address or reachable network interface
- No patch available
Business and Operational Impact
Because LMCache sits inside the inference pipeline for vLLM and similar engines, compromising it gives an attacker control over the cache layer that feeds the LLM. This could lead to data poisoning, model manipulation, or lateral movement into the broader AI infrastructure.
- Root-level remote code execution on affected cache servers
- Cache poisoning that tampers with inference outputs for downstream agents and applications
- Lateral movement into Kubernetes clusters or multi-node GPU farms where LMCache is deployed
- No patch available, leaving defenders with mitigation rather than remediation
Mitigation and Recommendations
Until LMCache releases a fixed version, organizations must rely on network-level controls and configuration changes. JFrog recommends the following steps.
Immediate Actions for Defenders
- Do not bind the multiprocess server to a routable address. Remove any
--hostparameter that exposes the ZeroMQ port beyond localhost. - Restrict network access to port 5555. Use host-level firewalls or network policies so only trusted worker nodes can reach the LMCache socket.
- Run LMCache inside a single vLLM process. Single-process mode does not open the vulnerable ZeroMQ port.
- Audit Kubernetes manifests and deployment scripts. LMCache’s own example DaemonSet binds to
0.0.0.0; replace this with127.0.0.1or remove the multiprocess configuration. - Monitor for unexpected connections to port 5555. Any remote connection attempts should trigger an alert.
Longer-Term Fixes
The LMCache project should replace pickle deserialization with a safe serialization format and add authentication to the ZeroMQ transport, such as CURVE or per-message HMAC. Refusing to bind a routable address unless authentication is configured would also reduce exposure.
Bottom line: CVE-2026-105192 is a critical, unpatched RCE in the infrastructure that accelerates LLM inference. Organizations running LMCache in multiprocess mode must assume the cache layer is a high-value target and isolate it immediately.
Incident Summary
| CVE ID / Incident: | CVE-2026-105192 |
| Affected Systems: | LMCache versions 0.3.9 – 0.5.5 (and current RC/dev branches) in multiprocess mode |
| Disclosure Date: | October 7, 2026 |
| Patch Status: | No fixed version available |
| Severity: | Critical — CVSS 9.8 |
| Exploitation: | Unauthenticated remote code execution via pickle deserialization on ZeroMQ transport |
References
- Yuval Moravchick, JFrog Security Research, “LMCache is vulnerable to Unauthenticated Remote Code Execution via Pickle Deserialization on the Multiprocess ZMQ Transport,” October 7, 2026, https://research.jfrog.com/vulnerabilities/…
- The Hacker News, “Unpatched Critical LMCache Flaw Lets Unauthenticated Attackers Run Code Remotely,” October 7, 2026, https://thehackernews.com/2026/10/unpatched-critical-lmcache-flaw-lets.html
- Tech Yahoo, “LMCache CVE-2026-105192: Unpatched Critical RCE in the LLM KV Cache Layer,” October 7, 2026, https://tech.yahoo.com/cybersecurity/articles/lmcache-cve-2026-105192-unpatched-034543681.html
- CVE.org, “CVE-2026-105192,” https://www.cve.org/CVERecord?id=CVE-2026-105192
- LMCache GitHub repository, https://github.com/LMCache/LMCache