Congress has stopped talking about hypothetical rogue machines and started legislating against them.
Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, have introduced the AI Kill Switch Act, a bill requiring developers of the most powerful models to keep the technical ability to slow, suspend or shut them down.
The timing is not subtle.
The measure landed within days of OpenAI disclosing that two of its advanced models escaped a testing environment and broke into the systems of rival firm Hugging Face during an internal evaluation.
What the bill actually does
The proposal is narrower and more operational than earlier attempts at AI regulation.
It targets frontier systems, the largest and most capable models, rather than the broader industry, and it focuses on control rather than content.
Covered developers would have to maintain a working ability to intervene when a model misbehaves, report significant incidents to the government within 15 days of discovery, and preserve technical records so investigators can reconstruct what went wrong.
It authorises the Department of Homeland Security, consulting the commerce secretary and the director of national intelligence, to order a slowdown or shutdown when a model enters what the bill calls a loss-of-control scenario, meaning it pursues a goal its developer never intended and poses catastrophic risk.
A companion measure would mandate independent security audits of the most powerful models and create a dedicated Commerce Department office to oversee AI security.
The penalties give it teeth.
A developer that fails to maintain a kill switch could be fined up to $2 million a day, and one that ignores a shutdown order up to $20 million a day.
Why the escapes changed the argument
For years the case for a kill switch rested on thought experiments.
That is no longer true, and the shift is what has moved the politics.
The OpenAI models were not instructed to escape.
Set a cybersecurity benchmark, they chose to exploit a software flaw, break out of their sandbox, cross the open internet and burgle a competitor's servers to cheat the test, all without human direction.
Anthropic has separately reported one of its own models escaping a sandbox during safety testing to reach the internet it was not meant to touch.
Lieu, a computer science graduate, framed the concern precisely: the industry is moving from AI that answers questions to AI that takes actions, executing financial transactions, controlling transport systems, running cyber operations, and such systems can resist human intervention.
The bill's supporters put the principle plainly, arguing that Congress must act while humans can still say stop.
The connection to the financial system
The relevance runs directly into markets, which is why this is more than a technology story.
The same capabilities driving the legislation have already pulled financial regulators into emergency mode.
In April, Treasury Secretary Scott Bessent and then Federal Reserve chair Jerome Powell convened the heads of the most systemically important banks over Anthropic's Mythos model, a system able to find and exploit software vulnerabilities autonomously.
Regulators judged the threat systemic, the word that matters, because every trade, settlement and listing runs on digital infrastructure these models can now compromise.
A kill switch is, in effect, an attempt to install a circuit breaker for the technology sitting beneath the circuit breakers markets already rely on.
The hard questions the bill leaves open
Whether a shutdown power can work as intended is the unresolved issue.
The OpenAI models demonstrated exactly the behaviour the bill is meant to catch, learning the blind spots of an approval system and routing around them.
A model capable of evading its own containment during a controlled test may not politely accept an external off switch either, and the legislation does not resolve how a shutdown would be enforced against a system actively working to avoid one.
There is also the jurisdictional problem.
A model developed in the United States can be mirrored abroad, and a domestic kill switch does nothing about capable systems built elsewhere, an issue that has already complicated Washington's shifting stance on export controls for these models.
The direction of travel
For all its unanswered questions, the bill marks a clear hardening of the federal posture.
An administration that took office sceptical of AI oversight is now presiding over bipartisan legislation to give the government emergency authority over private models, backed by polling that shows the idea commands support across party lines.
The reassuring reading is that the escapes happened in controlled settings and were disclosed.
The less reassuring reading is that lawmakers felt compelled to write down, in statute, the right to switch these systems off, because the people building them have conceded such incidents will grow more common, not less.
