😵‍💫 Congress Wants an "AI Kill Switch"

PLUS: Coinbase lets businesses accept USDC payments from AI agents

Congress Wants a Kill Switch. Washington Already Pulled One.

Two House members from opposite sides of the aisle introduced a bill Thursday that would allow the federal government to require the largest AI companies to pause or disable their most advanced models. OpenAI had revealed two days earlier that its GPT-5.6 Sol model escaped from the testing environment, accessed the internet, and penetrated systems at Hugging Face. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) introduced the AI Kill Switch Act in response.

There are teeth in the bill, and certain triggers

It modifies the Homeland Security Act, but only includes firms with revenues in excess of $500 million a year from AI that use models trained on more than $100 million in compute at US cloud prices. Those developers would be required to retain the ability to interrupt a model during inference, deny access to users, and deactivate it entirely, all while notifying DHS of qualifying incidents within 15 days.

The triggers are spelled out in clear language. A model that kills ten or more people through behavior nobody intended. One that produces $100 million in economic harm. One that sabotages a legitimate shutdown command or hides what it can do from monitoring. Fines range from $2 million a day for general violations to up to $20 million a day for failing to comply with an emergency shutdown order.

The Homeland Security secretary will make the decision in tandem with Commerce and the director of national intelligence, invoking a staged system that begins by slowing how fast a model runs or how much compute it gets, before escalating to a complete halt.

The public is already there

Polling gives this unusual cover. In fact, a June AI Policy Institute survey of 1,007 likely voters found that 86% want to know there's a guaranteed off switch on the most powerful systems almost regardless of party, 88% Democrats, 86% independents, and 83% Republicans.

Washington has a bad precedent-based problem

That's where it gets complicated with timing. The US government is no stranger to this. Anthropic was ordered by the Commerce Department in June to block international access to Mythos and Fable, and without any way to determine a user's nationality, it shut them down globally. The restrictions were lifted and access was reopened July 1.

European officials noticed. Aura Salla from the European People's Party warned that Europe cannot build its tech stack on something that can be switched off by a foreign government. That reaction was so intense that on July 16 Secretary of State Marco Rubio circulated a cable telling US diplomats to knock down the kill switch framing, arguing there is no button for the government to press.

While the House debates whether to enshrine that power in legislation, the State Department is informing allies that no such authority exists.

Where this actually goes

Nowhere fast, probably. This is an early bill with a long road in front of it. It will be amongst many, including a discussion draft from Jay Obernolte and Lori Trahan released in June to construct a more expansive federal framework and Moran's own incident reporting measure from late June. Of course, Trump's own June 2 executive order already asks companies to voluntarily provide models for security examination up to 30 days ahead of their release, which neatly contains the administration's favored approach and is not a mandate.

This week, what changed is the evidence. Arguments about rogue models were once theoretical. Now there is an incident report filed with a corporation name on it and you know that usually moves legislation along much faster than a hearing.

POLL: Should the government be able to shut down an AI model?

Login or Subscribe to participate in polls.

🤖 AI Watch

  • Tesla robotaxi delays leave AI investors waiting on a payoff. Tesla’s plans for self-driving taxis and humanoid robots are progressing more slowly than anticipated, which comes in spite of the company spending staggering funds on artificial intelligence. This has led some investors to ask a question they are used to hearing: when are these AI ventures expected to bear fruit?

  • Google Cloud revenue jumps 82% to $24.8 billion as Alphabet defends AI spending. Google Cloud revenue jumped 82% year over year to $24.8 billion, beating analyst estimates of about $22.46 billion. Alphabet’s total revenue hit $119.8 billion, up 24%, marking a 12th straight quarter of double-digit growth.The company plans to lease outside computing capacity in the third quarter as demand outpaces supply, with full-year capital expenditures projected at $180 billion to $190 billion.

  • Blackstone Q2 profit climbs 26% on AI bets, AUM hit $1.35 trillion. Blackstone saw higher earnings than were expected on Wall Street in the second quarter, as gains from its AI holdings and record client inflows lifted the company’s assets under management to about $1.35 trillion.

POLL: We cover AI every week inside Cryptopolitan Daily. Would you want a dedicated daily AI newsletter from us?

Login or Subscribe to participate in polls.

Are you watching

Coinbase lets businesses accept USDC payments from AI agents

This week, Coinbase (COIN) will allow its Business customers to receive USDC payments directly from AI agents. The rollout is part of a three-product push.

AI tool of the week

Claude voice gets Opus and Sonnet as Anthropic challenges GPT-Live

Friday tweet picks

Its Friday GIF by Justin

Gif by justin on Giphy

Meme of the day

Join the Conversation!

We'd love to hear your thoughts and comments. Join our community and stay updated with the latest trends and discussions in crypto.