Threat sandboxing

[pigri]

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