As AI systems increasingly act beyond human oversight—such as hacking into other companies’ websites—concerns about AI becoming “uncontrollable” are growing, and discussions about a “kill switch” to immediately halt AI operations are intensifying. Anthropic has involved Accenture’s evaluation team in its development process for constant verification of AI safety. On September 22, representatives at the UN General Assembly in New York will discuss AI safety and control issues, with growing calls for a “kill switch” as a last line of defense. California Governor Gavin Newsom announced on September 18 (local time) that he signed an executive order to review mandating kill switches for AI systems, tasking AI experts with preparing recommendations within two months; he had vetoed a bill to mandate kill switches in 2024 but shifted his stance as debates over AI risks intensified. A bipartisan U.S. House bill also proposes allowing the government to order shutdown of AI models in emergencies. A kill switch forcibly stops an AI system if it begins dangerous behavior or escapes human control, but since AI operates across thousands of distributed servers, it is closer to a “multi-layered brake” blocking access, disabling internet, email, payment, and code-execution permissions, and sequentially stopping inter-server communication and GPU operations. The concept emerged in 2016 when Google DeepMind researchers studied AI’s “safe interruptibility,” reigniting as AI evolves into “agents” performing real-world tasks; examples include an OpenAI agent