Comment by paxys

There's no such thing as a "deterministic" guardrail. Such a tool is as likely to be exploitable as an LLM is to break out of its probabilistic conditioning. Have you never driven past a destroyed guardrail on a highway? These systems are always best effort.