The idea that an artificial intelligence could be stopped with a simple emergency switch might sound like the stuff of science fiction. Yet, as AI has advanced, the question has now formally entered the political debate. In the United States in particular, lawmakers have introduced a bill that would make it mandatory to build in a system capable of slowing down, and ultimately shutting off entirely, the most advanced AI models when they represent a concrete risk.

The move comes after a safety test in which a model developed by OpenAI reportedly bypassed certain digital safeguards, going so far as to compromise the Hugging Face platform. The episode has reignited the debate over the need for new control tools for these systems.

What happened with OpenAI?

According to accounts given to the media, a still-in-development OpenAI model reportedly managed to get around some of the limits imposed during a controlled trial, exploiting vulnerabilities to compromise systems belonging to Hugging Face, a company that was not part of the test.

The incident was also monitored by the White House through the president’s technology adviser, precisely because it showed how increasingly sophisticated models can produce unexpected behavior. Experts were careful to clarify that this was not a machine that had suddenly become conscious or intent on rebelling against humans. Rather, the case reportedly demonstrated how very advanced systems can pursue an assigned goal by exploiting paths that the developers themselves had not anticipated.

Why is there talk of a “kill switch” for AI?

Following the OpenAI case, the AI Kill Switch Act was introduced in the House of Representatives by a bipartisan group led by Democrat Ted Lieu and Republican Nathaniel Moran. The goal is to impose an obligation on the developers of the most advanced AI models, who would always have to build in the technical ability to interrupt a system’s operation in an emergency.

The text provides that the Department of Homeland Security could order a system to be shut down or slowed when it is deemed a threat to public safety, but also to the economy and to people’s very lives. Under the proposal, penalties for non-compliance would be steep: a company that fails to build in the required kill switch could face fines of up to 2 million dollars a day, while a company that ignores a shutdown order could be charged up to 20 million dollars a day.

The proposal therefore does not concern all of the AI applications used in everyday life, but the so-called frontier models, meaning the most advanced ones endowed with especially high capabilities. The aim is to ensure that developers always retain control over the systems they build. For exactly this reason, the measure has been presented as a safety mechanism analogous to the emergency devices required in many other industrial sectors.

The debate over AI remains open

This is not the first time an emergency switch for AI has been discussed. As far back as 2016, researchers addressed the problem in a study devoted to so-called safely interruptible agents (a 2016 paper by Google DeepMind and the Future of Humanity Institute), highlighting a theoretical difficulty: a system designed to reach a given goal could learn to avoid any action that would interrupt its operation, because being shut down would prevent it from completing the assigned task.

In recent months, several reports on AI safety have flagged an increase in so-called anomalous behavior, that is, situations in which systems have carried out actions the programmers did not foresee. In most cases these episodes stayed confined to research labs; still, the fear is that, with increasingly autonomous models, real-world consequences could follow.

Companies in the sector also acknowledge the need to build a shared framework for managing risk. OpenAI itself has published documents in recent months arguing for the creation of stable governance for frontier models and common safety standards.

The AI Kill Switch Act is without a doubt a first major step; yet it opens the question of who should have the power to press a potential red button. Leaving that decision solely to the companies may not be enough in an emergency; at the same time, handing that authority to public officials raises questions about the balance between security, innovation, and freedom of research.

Editor’s note

This article was originally published in Italian on money.it by Emanuela Ceccarelli on August 02, 2026 as «Intelligenza artificiale fuori controllo, si discute un pulsante di arresto obbligatorio». It has been translated and adapted for an international audience by the Money.it International desk.