Can we keep AI from risking national security?

SALT LAKE CITY — Who wields ultimate power: the government or artificial intelligence?

This debate has lodged itself in the minds of many Americans from San Francisco to Washington, D.C., especially as AI becomes more powerful, and Utah Rep. Celeste Maloy is wading into the regulatory battle.

In a recent bill she sponsored alongside Rep. Sara Jacobs, D-Calif., Maloy is proposing that AI systems be tested in Department of Energy national labs for any nuclear-related national security risks before they're deployed publicly.

"Right now, we've got AI technology growing and developing very quickly," Maloy told the Deseret News. "And there are people in different parts of the government who are evaluating risks to highly classified information, like nuclear information."

Her new bill, the AI Threat Output and Monitoring Incident Containment Act (ATOMIC Act), would formalize a team that evaluates AI risks. Researchers would test AI for its jailbreaking techniques and develop contingency plans and mitigation strategies for large advanced AI developers.

The bill's text defines "advanced artificial intelligence" as AI trained on a quantity of computing power greater than 10 to the 26th floating-point operations per second.

"What we want to know is what capabilities AI has," Maloy said.

Andrew Caprio, a military legislative assistant working for Maloy, described the perceived need for the bill in a separate conversation with the Deseret News. "We can't let crisis make our decisions, so we need to always protect testing, evaluation and research," he said.

When asked whether he believed the bill could dampen AI innovation, Caprio said, "This is a relationship with large developers to then give another data set an outside look to find out what it is capable of and where it is going. That's not putting guardrails on it. Later on, that's a discussion the congresswoman and other lawmakers will have. But it all starts with research."

Maloy said, "the reason I'm doing this is because I have people approach me who really care that we get AI regulations right and not over-regulate it. We hope to be clear eyed about some of the risks and make sure that we're thinking ahead and not waiting until there's a problem," Maloy told the Deseret News.

What is an "AI nuclear incident"?

The bill's text defines an AI nuclear incident as:

• The generation of technical information, instructions or assistance that may likely serve to unlawfully develop, acquire or use a nuclear weapon or nuclear material.
• The generation of restricted data.
• The loss of control of nuclear weapons, material or facilities.
• A foreign terrorist organization or foreign adversary obtaining unauthorized access to, manipulating or interfering with such a system.
• Scheming behavior relating to weapons, materials, facilities or stockpiles.

Caprio described the research as "red teaming," which is a way to identify flaws, vulnerabilities, undesirable behavior or other dangerous capabilities AI may pose.

Which companies would need to comply with testing?

If passed, the law would apply to the largest AI developers. They would need to have invested $2 billion in AI development in the previous five years or have used an enormous amount of compute to train their models.

Current companies that have likely used that much compute include OpenAI (ChatGPT), Google DeepMind (Gemini), Anthropic (Claude), xAI (Grok) and Meta (MetaAI).

The bill promises to protect companies' proprietary business information and trade secrets.

When asked about concerns that smaller, less well-known AI companies may be better able to perform nefarious nuclear-related acts, Caprio referenced DeepSeek, a Chinese AI company.

"What DeepSeek did was ask Grok, Chat or Claude millions and millions of questions, basically wrote down their answers, then made predictions of what was the most likely outcome based on the question," Caprio said. "It doesn't actually process. It doesn't build. It can't take in new pieces that change every day."

In essence, AI companies that are worth more are the ones that have actually been taking in data, he said. Testing the hard hitters, Caprio predicted, would effectively test smaller companies, since many are trained on existing AI.

Related topics