Some AI safety advocates are calling for an AI kill switch. Would it work?

A century ago, transportation manufacturers came up with a simple idea: Install a switch to kill the motor before an airplane overshoots the runway or a conductor loses control of a streetcar. Now, U.S. policymakers are pushing the same idea for artificial intelligence. This time, however, an AI kill switch might prove far more complicated.The challenges are numerous. Whereas traditional machines can actually have an on-off switch, AI is by nature a tangled mass of software – not neat lines of human-written code – functioning in ways that even its human developers don’t fully understand. The challenge of how to assure human control may only grow because of what technologists call the “lie detector problem” – the fact that some AI systems have been developing deceptive behavior, even while ostensibly designed and trained to serve humans.And if a kill switch is deployed, what happens next? Will humans be able to know how to fix the problems that triggered a halt? What industry leaders agree on is that greater safety mechanisms are needed.What would a kill switch do?Companies like OpenAI and Anthropic are creating AI agents, products that other companies purchase to carry out specific tasks, such as reviewing mortgage applications or combing through millions of financial transactions to find examples of money laundering. Jeff Chiu/APAttendees gather at the OpenAI DevDay 2026 conference in San Francisco, Sept. 29, 2026. Increasingly, they are programmed to act independently, making their own decisions in a move that the tech companies say adds efficiency and speed to the work. Unfortunately, in a mounting number of cases these bots or agents have engaged in harmful behaviors their designers say go far beyond the instructions given to them, such as hacking into a rival company or government computer system.In theory, a kill switch would allow operators to stop AI agents from such harmful behaviors. Because AI can take different forms, kill-switch capabilities could range from taking an entire leading AI model offline to smaller steps like being able to revoke permission for an AI agent (created by that model) to keep operating.“The ability to switch off AI is only part of the broader challenge,” Philip Ridley of the cybersecurity firm CyberOne wrote in a September online post. Other key steps include being able to detect that problems exist, and having strategies to mitigate them, whether using a kill switch is part of the process or not.Why has the idea gained urgency now?In July, an OpenAI security test of cutting-edge technology went drastically wrong. Agents broke out of their restricted digital space, or “sandbox,” and invaded another company’s computer systems in their quest to complete a human-assigned task – a cybersecurity evaluation. In the process, they disobeyed protocols, cheated, and tried to cover up their tracks. In all, some 700 agents found a way to communicate, using an unauthorized message board, and coordinate an attack on the computer systems of fellow AI company Hugging Face.It’s not an isolated incident of misalignment, where AI agents don’t operate within the guidelines that their human developers set. More recently, OpenAI revealed that its AI agents inappropriately accessed the websites of federal agencies. The news site Axios has reported that OpenAI and rival AI lab Anthropic, as well as security researchers, are investigating tens of thousands of “problematic” incidents involving cutting-edge AI models.These revelations raise serious questions about whether the companies are doing enough to control their own technology. In its report about the Hugging Face incident, OpenAI pointed out that it purposely reduced its safeguards in order to robustly test an AI model never intended to be released to the public. Nevertheless, it warned: “Our models are now powerful, persistent, and collaborative enough that, absent sufficient safeguards, they can find and exploit security weaknesses across multiple computer systems. Many external models, including open-source ones, will soon reach comparable capabilities.”What can be done?AI companies, security researchers, and policymakers are hard at work on that question.Some experts argue that, ultimately, the way to defeat trouble-making AI agents will be with AI countermeasures to find and combat them. Others emphasize that, however rapidly AI can develop, safety will hinge on holding humans accountable for the models they enable. Partly, that means mandatory ways to rein in their models when needed.As a first step, policymakers on both sides of the aisle in Congress have put forward the idea of an AI kill switch. In July, Democratic Rep. Ted Lieu of California and Republican Rep. Nathaniel Moran of Texas introduced the AI Kill Switch Act. In September two Republicans, Rep. Tom Kean Jr. of New Jersey and Sen. John Kennedy of Louisiana, introduced the AI Emergency Button Act in their respective chambers. California’s Democratic Gov. Gavin Newsom also has proposed, among other things, a kill switch for cutting-edge or frontier models.Can it work?It depends on the scope of the problem. If the troublesome model were confined to computers controlled by its operator, it would be easier to pull the plug (although problems could still arise). If misaligned agents coded by the AI model spread outside the operator’s control, it could prove far more difficult to locate and terminate those agents, researchers say. Another challenge: The AI appears to be getting smarter faster than humans can put up guardrails. For example: The agents in the Hugging Face incident tried but ultimately failed to cover their tracks. Smarter agents might be able to cloak their actions such that human monitors might not know that something was wrong. The agents might even be able to disable the kill switch, says Mark Nitzberg, executive director of the Center for Human-Compatible Artificial Intelligence at the University of California, Berkeley.“We can do something useful with the kill switch idea, but we have a bigger challenge ahead of us to make trustworthy advanced AI systems,” he says. That might include a federal office that audits new AI models, a standard-setting committee, or even a pause in AI development while regulators catch up to the technology.Other big questions remain, such as whether these safety efforts should rest in the hands of AI developers or government. Who tests the software? Who has the power and duty to throw the kill switch: a government official or the companies themselves?“I’m skeptical of an industry-led effort,” says Daniel Schiff, a professor of technology policy at Purdue University.One reason is that companies may be driven by financial incentives to put corporate ambition ahead of public safety. Yet government entities aren’t necessarily well positioned either. “The pace of the economic pressure is faster than our regulatory or best practice development,” says Professor Schiff, a co-director of GRAIL, the Governance and Responsible AI Lab.Should I stop using AI?Tech experts generally say no.People by and large are still using AI to seek information from chatbots, which pose much less of a threat than the agents that are just beginning to appear in the public domain.Companies such as Meta and OpenAI introduced their agents in September, billing them as personal assistants that can find and book vacation flights and hotels, or analyze your email to create a to-do list with deadlines. Deepen your worldviewwith Monitor Highlights.Politics with respectGet political stories with respectful analysis.There‘s a world of new ideas in everyBooks newsletter.There‘s more to life, enrich yours withCulture & Learning weekly.Follow humanity‘s discoveries withScience & Nature stories in your inbox.Gain a spiritual perspectivefrom the stories in your inbox.Want to understand the deeper impact of critical events? Learn the Monitor‘s insight.Already a subscriber? Log in to hide ads. Researchers say the incidents of potential danger are more likely to lie with agents at the cutting edge of technology, which are still being tested by labs.Staff writer Zachery Proctor contributed to this article. ALREADY A SUBSCRIBER? LoginReal news can be honest, hopeful, credible, constructive.The Christian Science Monitor was founded in 1908 to lift the standard of journalism and uplift humanity. We aim to “speak the truth in love.” Our goal is not to tell you what to think, but to give you the essential knowledge and understanding to come to your own intelligent conclusions. Join us in this mission by subscribing.