high Cve 2026 105192 · AI Security

LMCache CVE-2026-105192: Unauthenticated Pickle RCE in Multiprocess Mode, Fixed in 0.5.6

Data graphic: LMCache one-message remote code execution flaw, CVE-2026-105192, rated CVSS 3.1 Critical 9.8, shown with a green stamp reading "Fixed in 0.5.6". Versions 0.3.9 to 0.5.5 are affected and the flaw needs no authentication; only routable binds are exposed, and workers and servers should be upgraded together.
AK

Threat intelligence editor · Published Oct 7, 2026 · Updated Oct 16, 2026, 8:57 PM EDT

CVE-2026-105192 is an unauthenticated pickle RCE in LMCache multiprocess mode, fixed in 0.5.6. Upgrade workers and servers together; routable binds were exposed.

What happened

Update (2026-10-16): LMCache has fixed this flaw in v0.5.6. The GitHub release was published on 14 October (22:21 UTC) and the package appeared on PyPI on 15 October. The release notes link GitHub security advisory GHSA-w5f2-9q9c-ff4h, which carries the same CVSS 3.1 score of 9.8. The fix is pull request #5644, merged to the dev branch on 12 October. It replaces pickle with versioned, typed MessagePack descriptors that are validated before use. The device wrapper type is chosen from a fixed local allowlist, there is no pickle fallback, and both the ZeroMQ and gRPC transport paths are covered. The wire format changed, so upgrade all workers and servers together. The affected range is unchanged: 0.3.9 through 0.5.5, the 0.5.6 release candidates the advisory lists (install final 0.5.6, not a pre-release), and dev builds before #5644. The project's example Kubernetes deployment bound all interfaces, and the advisory says deployments based on it, and on the official images that ran as root, "should be treated as exposed". 0.5.6 binds that example to localhost (PR #5684) and runs official images as a non-root user (PR #5685). We have not tested the patched release ourselves; this comes from the advisory, the release notes and LMCache's 13 October post. The sections below describe the flaw as it stood when we first published on 8 October 2026.

JFrog Security Research disclosed CVE-2026-105192 on 7 October 2026, a critical remote code execution flaw in LMCache, the KV-cache layer used alongside LLM inference servers such as vLLM (we covered the vLLM KV connector flaws). In multiprocess mode, LMCache opens a ZeroMQ ROUTER socket so worker processes can register and share KV cache blocks. That socket has no authentication, and incoming messages are decoded with Python's pickle.loads. A single crafted message is enough to run code as the LMCache process user.

NVD lists a CVSS 3.1 base score of 9.8 (vector CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H). The score comes from JFrog acting as the CNA; NIST had not added its own assessment, and the entry was marked Deferred when we checked on 8 October 2026.

Status when first published, 8 October 2026: we found no fixed release, and LMCache had not published a security advisory (The Hacker News said so, and the repository's GitHub security-advisories list was empty). The latest stable release then was v0.5.5 (12 September 2026), which JFrog lists as affected. That is no longer the case; see the update above. On 8 October we found no report of exploitation. Correction (2026-10-16): our first version of this article implied there was no public proof-of-concept. JFrog's write-up, published on 7 October, already contains a four-step proof-of-concept with Python code, and the GitHub advisory repeats it. We do not reproduce it here. NVD's SSVC data records Exploitation as "None" but Automatable as "Yes". CVE-2026-105192 was not in the CISA KEV feed when we checked on 8 October or again on 16 October, and on 16 October NVD still showed it as Deferred with SSVC Exploitation "None".

Why it matters

The practical risk depends on one setting: where the multiprocess server listens. JFrog and the CNA text in NVD both state that the default bind address is localhost on port 5555, and that the flaw becomes network-reachable when an operator passes a routable address with the --host flag. JFrog calls that "the documented multi-node setting", the way multi-node deployments let peers connect. JFrog also says LMCache used only inside a vLLM process does not open this port, and that the 9.8 score is for the routable configuration: a stock single-host install that leaves the default bind is not reachable from other machines.

Where that applies, an attacker with network access to the port needs no credentials and no user interaction. JFrog notes that the official container images run as root, so a successful attack on a containerised deployment would run as root inside the container. GPU inference hosts often hold model weights, API keys and cloud credentials, which raises the stakes of any code execution there.

This is the same class of bug that has recurred across AI serving stacks (we covered the LightLLM pickle RCE): ZeroMQ or similar sockets used for internal coordination, with pickle on the wire and an assumption that the network is trusted.

Technical details

According to JFrog, messages use msgpack, and extension code 1 is registered for DeviceIPCWrapper. When the decoder meets that extension code, it calls pickle.loads on the payload during argument decoding, before any validation. The socket offers no CURVE, ZAP or message authentication.

The LMCache source is consistent with this. In v0.5.5, the docstring of lmcache/v1/platform/base/ipc_wrapper.py says all wrappers share "the single msgspec ext code (1)" and describes "pickle-based (de)serialization". Unpickling attacker-controlled bytes lets the sender execute arbitrary Python, which is why NVD tags the issue CWE-502 (deserialization of untrusted data) and CWE-306 (missing authentication).

Attack chain: crafted msgpack message to the LMCache ZeroMQ socket, ext code 1, pickle.loads, code execution

How one unauthenticated message becomes code execution in LMCache multiprocess mode.

Affected versions: from v0.3.9 (released October 2025, per The Hacker News) through v0.5.5, the 0.5.6 pre-releases built before the fix (up to rc4; rc5 already contains #5644), and the dev branch as of 7 October. JFrog writes that "the decode path shipped in v0.3.9" and is still present in all of those. NVD's affected data lists 0.3.9 with an open upper bound, which agrees.

What defenders should do

  1. Upgrade to LMCache 0.5.6. Upgrade every worker and server in the same maintenance window, because old and new versions cannot talk to each other over the changed wire format. Rebuild container images from the 0.5.6 release rather than relying on an old tag. Steps 2 to 4 apply after upgrading too: the fix does not make the port safe to expose (see "Still not fixed after 0.5.6" below).
  2. Find multiprocess deployments. LMCache used only inside a vLLM process (the in-process connector) does not open this port, so those setups are not exposed to this CVE. Search for LMCache servers started in multiprocess mode and for anything listening on TCP 5555 (for example ss -ltnp | grep 5555 on the host or inside the container). Check launch scripts, Helm values and Kubernetes manifests for --host.
  3. Check what the bind address is. Review any manifest copied from LMCache's example Kubernetes deployment first: the version before 0.5.6 bound all interfaces. A localhost or loopback-only bind matches the default JFrog describes and is not network-exposed. Any routable address, 0.0.0.0 included, puts the socket on the network.
  4. Restrict reachability, before and after upgrading. Per JFrog, do not use --host with routable addresses and keep the port on localhost or trusted networks. Add firewall rules or Kubernetes NetworkPolicies limiting port 5555 to the specific worker pods. The Hacker News notes that firewalling lowers the risk but does not eliminate it, because any permitted peer can still send the message.
  5. Reduce blast radius. Do not run the server as root where you can avoid it, and limit the credentials and cloud roles available to the process.
  6. Hunt for abuse. The advisory says that if the port was ever reachable from an untrusted network, you should treat the host as potentially compromised: review connections to the port, check for unexpected processes or outbound connections from the LMCache process, rotate reachable credentials and redeploy from clean images. No indicators of compromise have been published. LMCache's post warns that exploitation happens before the message is processed, so a rejected or malformed request in your logs does not mean the attack failed. The advisory's reproduction notes that the server then logs a type error for REGISTER_KV_CACHE after the command has already run, so treat such errors from unexpected peers as a lead. Also look for unexpected connections to port 5555 from hosts that are not inference workers, and for child processes spawned by the LMCache process. JFrog's longer-term advice was to replace pickle with a safe format and add authentication (CURVE or HMAC). 0.5.6 does the first. LMCache's 13 October post says the RPC transport still requires no authentication, with RPC authentication planned for 0.5.7. It also says pickle remains on the engine-driven CPU-data / retrieve-response path, which it describes as disabled by default and not reachable from the network, with removal planned before 0.5.8. Fix PR #5644 is scoped to the device descriptors, so keep the port off untrusted networks even on 0.5.6.

Still not fixed after 0.5.6

LMCache says authentication for the RPC transport is planned for 0.5.7 and that the remaining pickle use on the engine-driven path will go before 0.5.8. Until then, a patched server is still not a server to expose to untrusted clients.

A separate LMCache CVE

NVD also published CVE-2026-107204 on 7 October, assigned by VulnCheck. The 0.5.6 release notes say the /run_script endpoint is now disabled by default and the admin HTTP server binds to localhost unless configured otherwise (PR #5129), but they do not mention that CVE, so we cannot say it is fixed. PR #5129 itself says the restricted builtins are not a security boundary, so turning the endpoint back on still means code execution; issue #5510 is still open. It describes a different bug: unauthenticated code execution through a /run_script HTTP endpoint in LMCache through 0.5.5, with a CVSS 4.0 score of 9.3. NVD shows the record as Deferred (last modified 8 October). It is a different component from the ZeroMQ flaw, so closing port 5555 does not address it. The NVD text says the bypass works by recovering real builtins through the injected FastAPI app object, and public GitHub issue #5510 is titled "run_script sandbox escape via FastAPI app object in LMCache internal API server (RCE when enabled)". We have not verified the details beyond that, so check whether any HTTP API server is enabled and reachable in your deployment.

What is still unclear

  • Whether the gRPC path could be exploited in practice. The text of fix PR #5644 says the gRPC device-wrapper codec also called the same deserializer through its pickled_payload bytes field, so changing only the ZeroMQ hook would have left another caller of the unsafe decoder. The gRPC transport (PR #4953) merged on 16 September, after 0.5.5 shipped, so it appears only in 0.5.6 pre-releases. We have seen no test showing it can be exploited.
  • Whether other reports are valid. Six GitHub issues (#5507 to #5512), filed around 6 October, allege unauthenticated access in other LMCache components, including a coordinator said to default to 0.0.0.0:9300. We have not verified them, and none has a CVE that we found.
  • Exploitation status. We have found no report of exploitation, but step-by-step reproduction code has been public since JFrog's 7 October disclosure, so treat unpatched routable deployments as urgent.

Sources

Keep reading

All latest →
  1. watchAI SecurityMicrosoft Shows Editing Fewer Than 10 Neurons Plants a Sleeper Backdoor in an AI Agent Model5 min
  2. elevatedAI SecurityMCPhantom: Coding Agents Keep Installing Archived MCP Servers That Will Never Be Patched6 min
  3. elevatedAI SecurityA Researcher's ARTEX Agent Escaped Its Sandbox and Breached WorkOS: What Leaked and Why It Worked7 min
  4. highAI Securityllama.cpp rpc-server Has Two Unauthenticated Flaws: A Stack Overflow Over RDMA and a One-Packet Crash4 min
  5. elevatedAI SecurityLoad/Mine: Fake Qwen2 Models on Hugging Face Turned GPU Inference Hosts Into Miners4 min
  6. elevatedAI SecurityClaude Code Before 2.1.251 Let Sandboxed Code Overwrite Files Outside the Workspace on macOS (CVE-2026-88968)3 min