MPs propose 'kill switch' to shut down rogue AI systems | Computer Weekly

"UK MPs are backing the introduction of an artificial intelligence (AI) "kill switch" that would give the government powers to shut down datacentres and computer systems in the event of an AI system risking a large-scale emergency."

"An amendment to the Cyber Security and Resilience Bill (CSRB), going through Parliament, aims to give the government powers to intervene if an AI system poses a "catastrophic risk" to people's lives, essential services or national security."

"Under the proposals, which have not been endorsed by government, the Secretary of State will be given emergency powers to turn off datacentres or AI systems if they are deemed to pose a significant risk. Risks could include adversarial uses of AI systems by state and non-state actors, the ability of autonomous AI systems to launch cyber attacks, and the development of "super-intelligent AI" that could escape human oversight."

"The proposals follow concerns around the development of frontier AI models, which are capable of finding previously unknown security vulnerabilities and developing ways to exploit them, posing risks to critical systems. Last month, Anthropic made its frontier AI model, Claude Mythos, available to selected technology companies, under Project Glasswing, following claims it discovered thousands of security vulnerabilities, including vulnerabilities that had been unnoticed for years."

An amendment to the Cyber Security and Resilience Bill would grant the Secretary of State emergency powers to intervene if an AI system poses a “catastrophic risk” to people’s lives, essential services, or national security. The proposal would allow shutting down datacentres or computer systems when a significant risk is identified. Supporters cite concerns that AI could undermine cyber security and disrupt critical supply chains if not properly regulated. Risks mentioned include adversarial use by state and non-state actors, autonomous AI launching cyber attacks, and “super-intelligent AI” escaping human oversight. The move follows worries about frontier AI models that can discover unknown vulnerabilities and develop exploitation methods, including claims tied to Anthropic’s Claude Mythos.

Read at ComputerWeekly.com