Kelly Unveils Framework to Protect Americans from Most Advanced AI Models   

“We should hold AI to the same standard we have held other dangerous technologies: the more dangerous it is, the tighter the controls.” 

Today, Arizona Senator Mark Kelly unveiled A Safety and Security Framework for Frontier AI, a plan to protect Americans from risks posed by the most advanced AI models. The framework draws on decades of experience and isbased on how America already treats risks like nuclear materials and weapons systems. It would establish clear safety standards for powerful AI models, deploy independent examiners to assess national security risks and evaluate companies’ safeguards as models are developed, and require rigorous testing and accountability for AI used in military operations. 

“I’ve spent my career around powerful technology: fighter jets, spacecraft, weapons systems. The rule has always been simple: the more dangerous it is, the more carefully you check it. AI should be no different. This framework makes sure the companies building the most powerful AI systems are held to the same standards we expect everywhere else when national security is on the line,” said Kelly. 

Kelly’s framework comes as whistleblowers from top frontier AI companies warn of the risks of unchecked acceleration of AI technology and following the Hugging Face attack where a powerful OpenAI model broke out of its controlled environment, exploited security vulnerabilities, and hacked into a separate AI company without being instructed to do so. He warns of the risks of advanced AI models being misused for cyberattacks or the development of weapons, all of which could harm Americans.  

The framework points to three problems: there is no common measuring stick for when a model becomes dangerous, companies oversee themselves with outside review only after the fact, and AI models used by the military lack standardized acquisition and deployment requirements. 

In the framework, Kelly lays out why the government has to step in: “The same advanced AI models that could help cure deadly diseases can also help someone develop biological weapons or launch devastating cyberattacks. Yet today, companies are deciding for themselves which models are dangerous and how much protection is enough. Too often companies act in the best interest of their bottom line, not the safety of the American people […] We can’t wait until Americans are harmed to act.” 

The framework answers with three pieces that build on each other: a rulebook, a referee, and accountability. Standards define the safeguards needed to address AI risks, independent examiners identify national security threats as models are developed, and new defense acquisition rules require rigorous safety testing and hold contractors accountable for hiding or neglecting known risks. 

Kelly also pushes back on the idea that the federal government shouldn’t step in to protect Americans: “Some will believe that the federal government can’t — or shouldn’t — take on this problem. They seem to have lost faith that we can do hard things. Looking at how we have handled other dangerous, rapid innovations in the past, one thing is clear: It’s not a question of if we have the power to do big, hard things and protect the American people — it’s if we have the courage.” 

Kelly’s framework draws on his experience as a Navy pilot and astronaut, and his national security oversight on the Senate Armed Services and Select Intelligence Committees. It also builds on Kelly’s leadership in AI. 

Last year, Kelly released AI for America, a roadmap outlining principles to make AI work for all Americans—not just large companies. In September, Kelly introduced the Make AI Work for Americans Act to require the companies benefitingmost from AI to pay into a federal fund that invests in American workers. Kelly has also introduced bills to improve research on how AI is changing the workforce, secure America’s AI systems from foreign adversaries, preserve human judgment and accountability as the military integrates AI with two bills included in this year’s Senate NDAA, study how AI affects older Americans, and protect seniors from deceptive AI chatbots. 

Click here to read the full framework.  

Print
Email
Share
Tweet