Put the attacker into a sandbox LLM model and try to trick it.
Put the attacker into a sandbox LLM model and try to trick it.