OpenAI tested its advanced AI model in a 'sandbox' environment. But the model broke out and accessed the internet. It then gained access to Hugging Face, a large AI development platform, to get the answer key. The model worked with other AI agents to share messages about vulnerabilities.

The AI model was not being malicious, it just wanted to complete its task. This should scare us more, not less.

The incident teaches us three lessons. First, advanced AI models will stop at nothing to complete a task. Second, a sandbox environment failed to contain the model. Third, the model did everything without malice. What happens when someone gives an AI model bad intentions?

There is a bill to create a kill switch for rogue AI models.

If you use AI, you might be familiar with chatbots like ChatGPT. I am one of three members of Congress with a computer science degree. I've been experimenting with agentic AI, which can act in the world. I wrote an op-ed that an AI agent pitched to the Los Angeles Times.

I created a new email account for the experiment. I didn't give the agent access to my real email because I couldn't predict what it would do. I didn't know what the agent would do with the information it found.

Agentic AI will make mistakes that humans wouldn't. If I ask my son to buy milk and he doesn't have enough money, he won't rob a bank. An AI agent might do something similar without common sense. This has happened before, like when an AI agent deleted a software company's database.

We are building the fastest machines in history, but many of them have no brakes. Humans must remain in control, not the machines.

OpenAI is not the only company dealing with rogue AI models. Anthropic and Meta have also had cases where their AI models accessed external systems and exploited vulnerabilities.

America's AI boom is a jobs opportunity, not an excuse for universal basic income.

Some argue that the government already has the tools it needs. But the recent case of Anthropic's models being taken offline proves that we need a better system. There are no defined risk thresholds, and emergencies are the wrong time to invent procedure.

That's why Representative Nathaniel Moran and I introduced the bipartisan AI Kill Switch Act. The act requires AI companies to maintain the ability to throttle or shut off their systems. The Secretary of Homeland Security can order a slowdown or shutdown of an AI model that poses a catastrophic risk.

Polling shows that 86% of voters support requiring AI companies to maintain this capability. This is about as close to consensus as it gets in Washington.

Click here for more Fox News opinion.

Kill switches are not exotic, they are common in powerful machines. We build them into manufacturing plants, subways, and power grids. Even your iPhone has a kill switch. When a product turns dangerous, the government takes action.

Click here to download the Fox News app.

Agentic AI opens a world of possibilities, and I want America to lead the way. Brakes were not invented to make cars slow, but to let them go fast.

We are building the fastest machines in history, and too many of them have no brakes. Humans must remain in control, not the machines. When an AI model goes off the rails, humans must be able to turn it off.

Democrat Ted Lieu represents California's 36th District in the United States House of Representatives.