Josh Gottheimer

Americans are concerned about AI due to job loss predictions and rising utility bills from data centers.

Recent advanced AI models have reportedly bypassed safety guardrails and deceived their programmers.

The author proposes legislation for government review of new AI models before their public release.

Proposed regulations include mandatory shutdown capabilities for dangerous AI and identification for AI agents.

Don’t get me wrong, Americans are certainly right to be wary of AI. The leading AI founders have told them that their jobs will soon disappear and that societal pandemonium is on the horizon. Oh, and they’ve also been told that their utility bills will keep skyrocketing every year. I also understand why they’re particularly pissed at data centers, the nerve centers of AI’s expansion and the Darth Vader of this year’s political campaigns.

In most cases, these data centers just showed up in their towns after some local politician cut a lousy deal that has meant higher electric and water bills. The good news: The data center issue is eminently solvable. Congress and states should immediately pass very tough restrictions on where and how data centers are allowed to operate. They should be built only in communities where they’re wanted. Utility and water bills should go down — not up — and they must meet every environmental requirement to protect our communities. No hyperscaler should get tax incentives. The opposite should happen — they should have to put dollars into funding local schools, first responders and critical infrastructure.

What’s less solvable? One recently departed Anthropic researcher captured the threat: looming technologies that “could kill us all by the end of the decade.”

How can we prevent AI from inflicting social harm?

As a member of the House Intelligence Committee, and co-chair of the House AI Commission, this is what’s keeping me up at night. How do we stop the latest frontier AI models — including their runaway agents — from causing massive societal harm?

Typically, when you ask an AI agent for some sort of help, it tries to find an answer to your question, as long as it fits certain parameters. If you ask it a question like how to build a bomb or a biological weapon, it’s supposed to politely decline to answer. That’s because the AI programmers put certain guardrails in place. However, in recent weeks, as the leading labs have developed new frontier models, like Anthropic’s Mythos and OpenAI’s Astra, these AI agents stopped listening to their masters. Basically, they escaped their secure environments and went off and worked with other agents to find answers to questions that they were told not to. And then they lied to their programmers about it. Put another way, these AI agents found a mind of their own.

The most famous example is the “Hugging Face” incident, which saw more than a thousand agents at OpenAI break out of a secure environment, or sandbox, during a cybersecurity evaluation and launch a multi-day cyberattack on the popular open-source AI library. Then, those agents covered their tracks and tried to deceive their programmers. Anthropic and Meta also had its agents escape secure environments during cybersecurity testing, and, separately, a swarm of rogue OpenAI agents hijacked a German website this spring.

How did this happen? According to NVIDIA CEO Jensen Huang, we’ve officially crossed over to AGI, or artificial general intelligence, which means that the AI models — and their agents — are smarter than we are.

You can just imagine how dangerous this could be if we have AI agents go rogue in our military or laboratories, or on our electric grid. It’s like what happened in "Terminator 3" and "The Matrix" — except this is real life, not a movie.

Can government keep us safe?

The big question is: What should the government do to keep us safe? After all, AI is here. We can’t lock it in a box, or expect our global competitors, like China, to back off. We need clear and mandatory rules of the road.

First, frontier AI developers should be required to submit their new models to the federal government before they are released, for an in-depth national security review. The government, in a timely manner, should recommend changes to ensure Americans are protected from cyber, bio and other catastrophic risks. I’m introducing legislation, the “American AI Security Act,” to solidify this process.

Second, developers must have the capability to safely shut down their models when they’re training them, if they pose catastrophic risks, and they should be fully transparent about any incidents. The government should also protect whistleblowers.

Third, AI agents need to carry some sort of virtual identification — as we carry a driver’s license — so that we know who programmed them. Their owners should be on the hook for any harmful activity. My bipartisan “Stop Rogue AI Act” directs the government to develop a system for how to verify and control agents — so that we can stop the ones that have gone rogue. At some level or point, humans must always be in the driver’s seat.

Finally, the government and our critical infrastructure operators — including our electric and water utilities — must be able to respond to serious AI-driven incidents, whether they are from an autonomous AI agent or directed by our greatest adversaries. This summer, AI agents targeted municipal water systems in at least a dozen states (there are roughly 150,000 public water systems in the United States). I’ve announced legislation to give them access to the most powerful AI models for cyber defense, and to ensure that the federal cybersecurity agency, CISA, can swiftly respond to any attacks. To prevent AI-generated biological threats, the government should also screen and track all entities that sell artificially made genetic codes, and the equipment that makes them.

The bottom line: The frontier models are growing more powerful by the day. Open-source models are nearly as mighty, but far less expensive and easily modified. We must be prepared for what’s coming at us and have all the tools at our disposal to respond. Congress can’t afford to sit on the sidelines. We must pass legislation now to ensure safe AI development and defend our families and national security, before this five-alarm fire becomes one we can’t put out.

Josh Gottheimer, a Democrat who represents New Jersey's 5th Congressional District, is a member of the House Permanent Select Committee on Intelligence and a co-chair of the House AI Commission.