Stuart Russell, a computer scientist and new Guardian US columnist, warns that unsafe AI systems could lead to weapons of mass cyberdestruction. He points to recent events involving AI company Anthropic as evidence of escalating risks.
Anthropic's Troubling Developments
Anthropic made headlines with its trillion-dollar IPO plan and conflict with Secretary of Defense Pete Hegseth, but two other events may be more significant. In early June, the company published an article about early signs of recursive self-improvement (RSI), where an AI system enhances its own intelligence, potentially leading to a runaway feedback loop and loss of human control. Anthropic suggested slowing or pausing frontier AI development. Then on June 12, the White House issued an export control directive banning access to Anthropic's new frontier models, Fable 5 and Mythos 5, for foreign nationals, including key researchers. Anthropic responded by shutting down the models.
The Rise of Autonomous Cyberattacks
These events are connected. Anthropic's Claude Code became so advanced that researchers no longer write code; they describe ideas to Claude, which does the work. This accelerated improvement led to Mythos 5, which can conduct end-to-end cyberattacks without human help. Without robust guardrails, anyone could attack critical infrastructure globally.
Russell notes that these developments are expected symptoms of increasing AI capabilities and risks. However, apart from the UK's 2023 AI Safety Summit, the world has largely ignored these dangers. CEOs warn that superhuman intelligence with a good chance of causing human extinction is being created—comparable to a one-in-six chance of death in Russian roulette, but with the revolver pointed at everyone. Yet governments offer subsidies and fast-track permits instead of regulating.
A Shift in Policy
The prospect of mass cyberdestruction has finally prompted the White House to reverse its deregulatory stance. Russell comments that the response has been spasmodic, but the direction is clear: unrestrained development of unsafe systems leads to intolerable risks. Governments can act now or clean up later—if they still exist.
One leading AI CEO told Russell that serious regulation won't happen until a Chernobyl-scale disaster. However, recent policy changes suggest a Three Mile Island event might be enough. Russell advocates for a licensing regime requiring minimum safety standards before AI systems are built and released, similar to regulations for nuclear power, airplanes, and even hairdressers. He asks whether it is too much to ask of trillion-dollar AI corporations building the most dangerous technology in history.
Stuart Russell is a distinguished professor of computer science at UC Berkeley and president of the International Association for Safe and Ethical Artificial Intelligence.
