Local code execution via caller-evaluated argument

Code injection in Gradio 4.36.1 was reported to let local users execute Python through a crafted create_pyi argument. The proof of concept evaluates exec() while Python prepares the class_code argument, before create_pyi passes the resulting value into a Jinja template; the function does not evaluate that input. GitHub withdrew the advisory because the caller must already run the crafted Python in the same process, so no separate attacker-controlled trust boundary or additional capability is demonstrated.