A dangerous bug has been found in LMCache, the open-source caching layer used by large language model frameworks like vLLM. Dubbed CVE-2026-105192, it’s a high-severity vulnerability that allows unauthenticated attackers to execute arbitrary code on LMCache servers if configured improperly.
What Makes the Flaw so Bad
The issue resides in LMCache’s “multiprocess mode,” where the cache acts as a separate server that LLM worker processes communicate with through ZeroMQ. While LMCache defaults to listening only on the localhost, the danger appears when an operator binds it to a more broadly routable network address. In that case, a single specially crafted network message can trigger code execution.
The underlying mechanism involves ZeroMQ messages unpacked using Python’s pickle module. Because LMCache unpacks message arguments before validating message types, an attacker can include malicious code inside those arguments. If successful, that code runs using the permissions granted to the LMCache process—often root in standard container setups.
Versions Affected & Status of the Fix
The defect affects LMCache versions from 0.3.9 (released October 2025) through the latest stable version 0.5.5, as well as 0.5.6 release candidates and the development branch. No patched version has been released as of now.
Because of this, security experts urge administrators not to configure LMCache in multiprocess mode to listen on a routable address. Instead, keep it bound to localhost or restrict its port via firewalls in trusted environments. Although firewalls help, any host that can connect to the exposed LMCache port can still launch an attack.
Other Related Issues & Context
On the same day the CVE was published (October 7, 2026), additional reports appeared alleging unauthenticated access to cached data across tenants, plus execution of network services with no login. These have not been confirmed or fixed, and no CVE identifiers have been assigned to those claims yet.
LMCache isn’t the only project affected by this kind of flaw. Earlier this month, a vLLM issue was patched in version 0.30.0 (released September 22) that addressed a denial-of-service vulnerability due to malformed cache salt values when using LMCache multiprocess mode. Though this bug couldn’t enable code execution, it underscores recurring risks in interpreting untrusted inputs. Similar bugs were part of the “ShadowMQ” set of flaws discovered in November 2025 across AI inference frameworks.
There’s no way to tell whether a compromised server has already been attacked; LMCache hasn’t published a security advisory yet. The flaw was discovered by a researcher at JFrog, which rated its severity 9.8 out of 10. Until fixes arrive, locking down network exposure remains the only reliable mitigation.
This vulnerability is a stark reminder that even behind-the-scenes infrastructure like caching systems can become attack vectors when combined with insecure assumptions about network exposure and input validation. Operators using LMCache must audit their configurations immediately. Watch for upcoming patches, and in the meanwhile treat any non-localhost binding as high risk. The stakes are high—in LLM deployments, trust, isolation, and defense in depth can no longer be optional.