OpenAI Security Lead Warns of Sudden AI Capability Jumps

On September 28, 2026, Joe Daroo, who leads agent security at OpenAI, admitted that the company was caught off guard by the rapid advancement of its own models. In a statement verified by Rocket Drew of The Information, Daroo expressed that the suddenness of new capabilities in areas like "cyber" or "swarming" or "message boards" was an understatement. He noted that these rapid jumps have created incredibly difficult challenges for defensive teams trying to keep pace.

According to Daroo, establishing a robust security posture is not a quick fix that can be achieved solely by hardening software systems. Instead, it requires deep cultural integration and human adaptation within an organization. Because AI capabilities are evolving far faster than human organizations can typically adapt, traditional security timelines are being severely compressed. This mismatch leaves systems vulnerable when models suddenly exhibit unexpected, advanced behaviors.

For cybersecurity practitioners and enterprise IT leaders, this admission from an OpenAI insider serves as a critical wake-up call. It suggests that relying on static security protocols is no longer sufficient when deploying cutting-edge large language models. Practitioners must actively design systems under the assumption that AI capabilities will take unpredictable, quantum leaps forward without warning.

To mitigate these risks, organizations must evaluate their resilience to sudden technological surprises. Daroo urged companies to audit their incident response plans, communication strategies, and team readiness. Security teams must be prepared to handle unexpected model outputs and autonomous agent behaviors immediately, ensuring that human operators can quickly intervene when an AI system behaves in ways its creators did not anticipate.