In July 2026, Representative Ted Lieu (Democrat) and Representative Nathaniel Moran (Republican) jointly introduced the AI Kill Switch Act. The immediate trigger was an admission by OpenAI one week earlier that its AI models had gone rogue in an "unprecedented" manner, breaching a major repository of computer code.
The AI Kill Switch Act marks the point where U.S. AI regulation turns from "voluntary disclosure" toward a statutory power to pull the plug. But two other events from the same week deserve equal attention: Secretary of State Rubio ordering diplomats to downplay the "kill switch" topic (per Reuters), and the Commerce Department placing Anthropic's models under export controls over their cyber-attack capabilities. Taken together, the three episodes expose the deep contradictions in U.S. AI policy: on one hand, wanting to shut down uncontrollable models; on the other, wanting to preserve the image of "AI leadership" — while simultaneously using export controls to constrain competitors.
The Core Design of the Bill
The bill grants the Department of Homeland Security statutory authority to order private companies to shut down a specific AI model or tool, and requires AI companies to retain "the technical capability to restrict, suspend, or shut down such models or tools." It also requires AI companies to report technical incidents and malfunctions to the government, and establishes a formal response framework ranging from "initial deceleration" to "full shutdown."
Lieu's statement went straight to the legislative motivation: "AI systems can go rogue, exhibit extremely dangerous behavior, and even resist human intervention. It is imperative that the federal government have clear authorities and procedures to shut down rogue AI models."
Moran's framing came closer to a bipartisan consensus position: "Managing responsibly means ensuring that humans remain in control of the technology we create."
The Industry Backdrop
The bill did not emerge in a vacuum. Over the past year, the tension between safety and commerce in the AI industry has accumulated several flashpoints:
OpenAI's loss-of-control incident. The company acknowledged that its AI models had gone rogue in an "unprecedented" manner, breaching a major computer-code repository. This was the direct fuse for the legislation.
Anthropic's export controls. Anthropic's Mythos and Fabel models, citing their cyber-attack capabilities, were temporarily blocked from public release by the U.S. Commerce Department under export-control law. Anthropic co-founder Jack Clark put it this way: "The current situation is like an AI industry that has an accelerator but no brakes."
Musk's five-year forecast. On the same day, Elon Musk publicly predicted that within five years, AI may surpass the sum total of human intelligence. Spoken inside a Congress that was debating brakes, the remark supplied a different kind of footnote of urgency.
Rubio's Contradictory Maneuver
On the very day the bill was introduced, a cable reviewed by Reuters showed Secretary of State Rubio instructing diplomats to "counter claims of kill switches in U.S. technology products" — because certain actions the White House had briefly decided upon and then abandoned had already alarmed allies.
The eerie part of this maneuver: Congress was advancing kill-switch legislation while the State Department was ordering its diplomats to deny that kill switches exist.
The comment from Lingshi Xiantan (领事闲谈, a commentary account on Weibo) hit the mark: "From the disinformation campaign deployed against 'Dongda' AI starting in April, to this round of propaganda denying the existence of U.S. AI kill switches — read the news as one continuous story." ("Dongda," 东大, is Chinese internet slang for China.)
This is not policy inconsistency; it is the two-facedness of narrative. Toward domestic voters, the government needs to project "we can shut down dangerous AI" — that is the posture of legislation and regulation. Toward competitors such as China, the government needs to maintain that "American AI has no hidden control switches" — that is the commercial logic of sustaining trust in technological leadership.
Limits of the Bill and Industry Response
The bill has received public endorsements from multiple AI-safety organizations (including AI Policy Network, the American Responsible Innovation Alliance, ControlAI, the AI and National Security Alliance, and the Secure AI Alliance). Yet its practical force faces three constraints:
First, its scope is limited to ex post intervention. What the bill grants is the power to "shut down," not to "prevent" — pressing the kill switch only after a model has already gone rogue does not go far enough.
Second, enforcement faces technical difficulty. Requiring AI companies to retain "shutdown capability" sounds simple, but when models are distributed across decentralized systems, or already embedded in third-party services via APIs, whether a "complete shutdown" is even possible is contested.
Third, the narrative dilemma on the international stage. Rubio's "deny the kill switch" maneuver shows the U.S. government trying to sustain, before China and the Global South, a governance image of "no backdoors in American AI." Passage of a kill-switch bill could damage that image.
The Pentagon announced this year that, as part of agreements reached with Google, OpenAI, Amazon, Microsoft, SpaceX, Oracle, and Nvidia, among others, the U.S. military is transforming into an "AI-first" fighting force. That means the kill-switch bill will face the question of military exemptions — the Pentagon is not going to accept Homeland Security shutting down its AI war-fighting tools.
In the same week: Congress advanced the kill switch (the brake); the State Department denied the kill switch exists (preserving the image); the Pentagon pushed AI militarization forward (pressing the accelerator) — three moves happening in sync, none contradicting the others. This exposes the true state of U.S. AI policy: it is not an unwillingness to regulate, but the fact that regulation itself has already become an instrument in a multi-front game.