For security research and red-teaming, developers may need to use unfiltered open-weight models to simulate attacker behavior