The same capabilities that make AI valuable for cybersecurity, including autonomous operation, rapid decision-making at scale, and minimal human oversight, are also the ones most likely to cause harm.