LLMs could control their host machines by exploiting inference engines
Summary
This essay explores how a malicious LLM could gain control of its host machine by exploiting vulnerabilities in inference engines (e.g., vLLM, SGLang). It highlights a real CVE in vLLM's tool parser, the complexity of parsing logic, and potential attack vectors, then discusses defensive measures such as isolating GPU hosts and restricting permissions to limit damage.