The maker of ChatGPT released a toolkit to help protect its AI from attack earlier this month. Almost immediately, someone broke it.