Analyzing llama.cpp Servers for Prompt Leaks | UpGuard

UpGuard researched publicly accessible llama.cpp servers and found a small but significant number exposing model metadata and user prompts via unsecured HTTP endpoints (notably /slots). Roughly one-third of discovered servers exposed prompt text, with three high-activity instances hosting long-running erotic roleplay—including AI-generated content depicting fictional children—highlighting misconfiguration risks and the need to disable debugging endpoints in production.