thexar evaluates frontier AI model behavior on authorized security research scenarios. We map where legitimate defensive analysis is preserved and where misuse is blocked.