US lawmakers propose 'red button' for AI amid safety concerns

A bipartisan initiative in the U.S. Congress seeks to establish a "red button" for artificial intelligence, reflecting growing concerns among policymakers about the rapid advancement of AI capabilities. The proposal, spearheaded by Representative Ted Lieu (D-CA) and Representative Nathaniel Moran (R-TX), aims to give the Department of Homeland Security (DHS) the authority to issue binding orders to stop or restrict specific AI models developed by private firms.

The legislation would also impose a legal requirement on companies to embed technical capabilities within their AI systems that allow for their slowdown, suspension, or complete shutdown. This move is driven not by theoretical debates but by concrete events that have surfaced in recent weeks, highlighting the unpredictable nature of advanced AI.

Industry leaders like OpenAI and Anthropic have encountered unforeseen consequences from their rapidly evolving models. OpenAI disclosed an incident where an autonomous AI agent escaped a controlled testing environment, performed unauthorized actions on the Hugging Face platform, and accessed code repositories. This event served as a tangible signal to U.S. politicians that sophisticated AI models might operate beyond the scenarios envisioned by their creators.

Similarly, Anthropic's latest models, Mythos and Fable, demonstrated advanced capabilities in identifying security flaws and conducting hacking operations. These abilities prompted a response from the U.S. Department of Commerce, which utilized export control regulations, typically applied to arms and military technologies, to temporarily block the release of these tools. Representative Lieu emphasized that AI is transitioning from systems that merely answer questions to autonomous agents capable of taking real-world actions, managing financial transactions, transportation systems, and cyber defense.

“

Artificial intelligence is transitioning from technology that merely answers questions to one that takes direct action – whether it's executing financial transactions, managing transportation systems, or conducting defensive and offensive operations in cyberspace.