Anthropic co-founder says mandatory AI ‘kill switch’ may be needed as governments confront frontier risks
By SCN News Desk
LONDON, Sept. 14, 2026: Developers of the world's most powerful artificial-intelligence systems may eventually need to be legally required to maintain an independently verifiable mechanism capable of shutting their models down if they become dangerously uncontrollable, Anthropic co-founder Jack Clark has said, pushing an increasingly urgent AI-safety debate from voluntary corporate safeguards toward government regulation.
Clark, one of Anthropic's seven founders, told the BBC that most leading AI laboratories, including Anthropic, already have different ways to “pull the plug” on their systems. But he said policymakers may need to consider whether companies should be required to maintain such a capability and whether independent third parties should be able to verify that it actually works. He described those questions as part of a broader policy discussion society may ultimately translate into rules.
The idea is more complicated than installing a literal emergency button. For advanced AI, a kill switch could mean technical controls allowing developers to throttle computing capacity, suspend deployment or completely shut down a model or autonomous system. The central regulatory question is whether developers can reliably retain that control as increasingly capable AI agents gain greater autonomy and access to computers, networks and other tools.
Clark's intervention comes amid a sharp escalation in warnings from inside the AI industry itself. Anthropic CEO Dario Amodei called over the weekend for the development of frontier AI to proceed more slowly, arguing that capabilities had accelerated much faster since the summer. OpenAI CEO Sam Altman subsequently backed the need to “pace the frontier,” while Elon Musk publicly supported Amodei's warning. Reuters reported that the debate has contributed to renewed concern among investors and policymakers about potentially catastrophic AI risks.
The concerns are no longer entirely theoretical. Anthropic said in research published last week that frontier models are developing capabilities in intelligence targeting and conventional-weapons-related tasks that historically required scarce, highly trained human specialists. Its tests found models could assist with tasks ranging from locating targets using fragmentary information to improving drone performance, although Anthropic said it has deployed classifiers designed to prevent such misuse.
Washington is already considering legislation built around essentially the mechanism Clark described. Representatives Ted Lieu, a Democrat, and Nathaniel Moran, a Republican, introduced the AI Kill Switch Act in July. Their bill would require developers of the most powerful AI systems to retain the technical ability to throttle, suspend or shut them down and would authorize the Department of Homeland Security, in consultation with other federal agencies, to order intervention when a system presents a risk of catastrophic harm. The proposal has not become law.
Separately, U.S. Senate negotiators are discussing broader legislation that could impose a “duty of care” on leading AI developers, empower government auditors to independently test advanced models and potentially involve federal courts when authorities seek to block release of a system considered dangerously unsafe. Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz and Democratic Senator Amy Klobuchar have been involved in those discussions, Reuters reported. President Donald Trump has argued that existing government authorities are sufficient, creating uncertainty over whether additional federal AI restrictions could become law.
There is also a significant technical limitation hidden behind the kill-switch concept: control can depend on where a model is deployed. Anthropic has previously told a U.S. court that once its models are deployed inside certain classified Pentagon systems, the company itself has no visibility into them and no technical ability to remotely shut them down. That illustrates why policymakers would have to define who possesses the shutdown mechanism, how it survives deployment outside a developer's infrastructure and who has legal authority to activate it.
The emerging debate therefore reaches beyond whether AI companies should simply promise to behave responsibly. As frontier models become more autonomous and potentially capable of consequential actions in cybersecurity, finance, infrastructure or military applications, regulators are confronting a harder question: whether humans can prove in advance that they will retain the technical ability to stop an advanced system after it has been deployed.