Hacks & Data Breaches I

LMCache Vulnerability

Heads up, everyone using LMCache to speed up their AI models, there's a security issue you should know about.

Problem is there's a critical flaw (CVE-2026-105192) that lets hackers run code on your server remotely without a password. It's rated a 9.8 out of 10 on the severity scale. A second issue (CVE-2026-107206) also lets anyone get in and mess with your settings or delete data. This is mainly a problem if you're running LMCache (versions 0.3.9 to 0.5.5) and have it set up to talk to other computers over the network. If it's only running on your local machine, you're likely okay.

The best move to take is to not let LMCache listen on an open network address. Make sure it's only accessible from your own machine or a trusted, private network.
Use a firewall to block all outside access to the LMCache port.

There isn't a fix available yet, so keep an eye out for an official patch and update as soon as it's released.
Basically, lock down your LMCache server until the developers release a fix. It's better to be safe than sorry!

↑ 2

Debate

Comentarios

2

Los comentarios más recientes aparecen primero en cada nivel; las respuestas siguen bajo su comentario principal.

Escribir Únete para participar en la discusión.

Firewalling the port is only half of it - LMCache chats with a separate model server, so if that link is also exposed, the mitigation chain has a hole. Push both onto the same private interface rather than trusting one rule.
↑ 0 Responder
bind to localhost or a private interface is the whole mitigation here, and it's cheap to do. worth noting the 0.3.9 cutoff too - if you're on something older than that, unclear whether you're affected or just uncharted, so check the advisory rather than assuming you're in the clear.
↑ 0 Responder