Can the U.S. Trust AI With National Security?
Nowhere are the stakes higher for making sure the systems stay aligned with their creators’ purposes.
By Judd Rosenblatt and Cameron Berg
Richard MiaWhen nuclear strategist Herman Kahn published “Thinking About the Unthinkable” in 1962, it compelled Americans to confront the possibility of nuclear war and civilizational collapse.
Military and political leaders initially dismissed his work as alarmist.
But Kahn’s core insight was crucial and simple: Hoping that things won’t get completely out of control isn’t a plan.
We face a similar inflection point with artificial intelligence.
We’re racing toward systems with capabilities that seemed like science fiction a year ago.
The Pentagon is negotiating contracts with AI labs, each with ceilings up to $200 million, to integrate AI into national security.
But we’re buying raw AI firepower without any comparable investment in making these systems steerable and secure.
Recent research has demonstrated that AI models can harbor sleeper agents that can be triggered under specific conditions without detection.
The same adversaries that compromised U.S. telecommunications through Salt Typhoon and targeted critical infrastructure through Volt Typhoon can exploit vulnerabilities in military AI systems.
We know that Chinese intelligence has penetrated frontier AI labs: In 2024, a former Google engineer was indicted on charges that he stole AI trade secrets while secretly working for Beijing.
(He has pleaded not guilty and his trial is scheduled to begin next month.)
The way to address this is with AI alignment research—ensuring that systems’ objectives and reasoning stay stable, predictable and faithful to their intended mission across new situations, long time horizons and adversarial pressures.
The administration’s newly announced Genesis Mission at the Department of Energy launches a coordinated national effort to accelerate AI-enabled scientific discovery.
These are appropriate investments in cutting-edge capability, but a failure to invest comparably in alignment can allow adversaries to compromise American AI capabilities.
While American defense procurement treats alignment as an afterthought, China is moving systematically.
In January, Beijing launched an $8.2 billion National AI Industry Investment Fund.
Since the 2017 New Generation AI Development Plan, Chinese policy has emphasized building AI that is “safe, reliable and controllable.”
Chinese military strategists stress the importance of systems that remain under operational command.
Beijing grasps what many American policymakers miss: Alignment research accelerates AI capabilities.
Whoever solves these technical problems builds more capable systems and wins the race.
Many AI policy discussions instinctively cast alignment as compliance overhead that slows development.
But historically, such constraints have often unlocked capabilities.
The F-16 became the most successful fighter in history not despite its strict design constraints, but because of them.
The same principle applies to AI. Giving models structured frameworks for how they think produces dramatic capability improvements, and techniques designed to make AI more aligned have been adopted by most major labs.
Yet most promising alignment directions remain unexplored.
The systems that defense planners need for extended autonomous operations require alignment properties we’ve barely begun to develop: stable long-term objectives, interpretable reasoning across complex decision chains, verified shutdown protocols that can’t be circumvented, and principled resistance to adversarial manipulation.
Military-grade reliability demands military-grade commitment to alignment research.
Unlike cybersecurity, where we’re locked in a cycle of patching vulnerabilities after adversaries exploit them, AI alignment offers a chance to build security from the ground up.
We can establish verified shutdown protocols, interpretable reasoning systems, and resistance to adversarial manipulation before these systems are deployed.
But this chance won’t last forever.
Frontier lab leaders expect systems that can match human experts across all cognitive domains within 12 to 18 months.
Once models can autonomously design successor models, and once powerful AI systems are embedded in critical infrastructure and military operations, we’ll face the same reactive posture that has plagued cybersecurity for decades.
Three steps would position America to win:
First, launch broad-scale programs that target neglected alignment research.
Private labs optimize for commercial performance, which means critical security challenges go unsolved.
These problems lack commercial importance, but they determine whether AI systems can be trusted with national-security operations.
The government must fund this research directly, just as it has historically funded cryptography, semiconductor security and other dual-use technologies for which market incentives misalign with defense needs.
Second, require military-grade alignment research and development in major Defense Department AI contracts.
The director of the Defense Advanced Research Projects Agency has already set a goal to achieve military-grade AI, and the Pentagon is already spending hundreds of millions on frontier capabilities.
Those contracts should mandate that computing resources go toward interpretability, verified shutdown protocols, the elimination of sleeper agents and long-term objective stability.
This approach would motivate private-sector research while ensuring that the systems we deploy meet defense-grade security standards.
Third, build special zones for AI research on federal land with alignment requirements built in from the start.
The Pentagon already demands the best equipment, training and strategic doctrine.
It should demand the same from AI systems: reliably steerable toward American strategic objectives across time horizons that matter.
America built the atomic bomb when the physics seemed impossible.
We reached the moon when the engineering seemed fantastical.
We created the internet when networked computing seemed impractical.
Americans have always risen to civilizational challenges when we’ve seen them clearly and moved with conviction.
The challenge now is to build AI systems we can trust with the future.
That future is closer than most realize, and the window for shaping it is open.
But it won’t stay open forever.
Mr. Rosenblatt is CEO and Mr. Berg a research director of AE Studio.
0 comments:
Publicar un comentario