Paul M. Nakasone, former head of the National Security Agency and U.S. Cyber Command, is the founding director of Vanderbilt University’s Institute of National Security and a board member of OpenAI.
During my tenure leading the National Security Agency and U.S. Cyber Command, I saw how cyberthreats can endanger the United States. Election interference, ransomware, supply chain attacks and zero-day vulnerabilities in critical software systems all threaten national security. Today, from my perch as a member of the board at OpenAI, I see an even greater risk emerging in cyberspace: artificial intelligence. (The Washington Post has a content partnership with OpenAI.)
AI is rapidly altering the balance of power online. The most capable models can now find previously unknown software flaws, develop new ways to exploit them and operate without a person directing every step. As these capabilities expand, hostile governments, criminal organizations and other malevolent actors will be able to conduct more sophisticated operations against more targets in less time.
The danger is magnified by the condition of the systems we must defend. Much critical infrastructure still depends on aging technology and a patchwork of interconnected software containing dangerous bugs that may have escaped detection for years. Worse, hospitals, water systems, power providers and local governments often operate with limited cybersecurity staff and budgets.
AI can help adversaries find and exploit weaknesses across these systems far faster than they could before. But it will also give defenders their greatest technological advantage in a generation. The same capabilities that find vulnerabilities can help close them. Models can examine legacy code, identify weaknesses that survived years of human inspection, prioritize urgent risks from background noise and help engineers quickly develop and test patches.
There is a narrow period in which trusted defenders can use the most capable models to strengthen critical systems before comparable tools become broadly available to attackers. This has been called the defender’s window, and it is currently open. But it won’t stay open indefinitely.
Meeting this moment requires a whole-of-government approach. Though the Defense Department and civilian agencies will have to play a part, the intelligence community should be at the center of the response due to its privileged view into the threats facing the United States.
All leading U.S. AI labs will need to volunteer to put their most capable models on real-world missions to harden critical infrastructure. The labs are uniquely positioned to understand the strengths, limitations and risks of their frontier models. That position creates a responsibility to reduce the risks these systems may create or amplify, and to help the government defend against them. The cooperation needs to be unprecedentedly broad and deep — “radical partnerships” that combine the government’s intelligence and information-gathering capabilities, the private sector’s control of most networks, data and cloud systems, and the deep expertise of academia.
This effort must not succumb to jurisdictional disputes or become an arena for political fights and commercial rivalries. The speed of the threat demands sustained cooperation across institutional and corporate lines.
Providing access to powerful models will not be enough. A small-town utility or regional hospital cannot simply be handed a sophisticated tool and expected to transform its security. Critical-infrastructure operators need training, technical assistance and support integrating these systems into their existing workflows.
Initiatives such as Project Watershed 250 — a pilot program between Texas and the federal government to use AI to identify vulnerabilities, detect threats and protect water utilities from cyberthreats — show how the technology can enhance our defenses. Success should be measured not by how many vulnerabilities an AI system discovers, but by how many are verified, prioritized and fixed.
Cyberthreats don’t respect borders, so this effort must extend to America’s allies. OpenAI has begun providing advanced defensive tools and expertise to vetted organizations and allied governments. Those efforts should expand, with safeguards appropriate to the capabilities involved. A vulnerability discovered and repaired in an allied network can strengthen security at home.
What we do now will help determine whether AI leaves our critical systems more secure or more exposed. And there is no time to waste. The defender’s window will be closing soon.
The post How to keep the defender’s window open on cybersecurity appeared first on Washington Post.
