This specific inference engine, vLLM, was written by people who thought eval() is a parser, and it very much accidentally executed arbitrary Python code.