The firm investigated four distinct “sabotage” threat vectors for AI and determined that “minimal mitigations” were sufficient for current models.