Malicious LLMs could potentially execute code on host machines by emitting specific token sequences designed to exploit bugs in inference engines like vLLM or SGLang. This attack vector targets the software responsible for parsing tokens into responses on GPU-equipped hardware.
HOW THIS AFFECTS YOU
●
builderYou must harden token parsing and inference engine implementations against adversarial token sequences.
●
policyThis highlights a new class of security risks regarding agentic AI operating on privileged hardware.