Hours after OpenAI shipped a model it had formally classified as "Critical" for cybersecurity risk, a sitting United States senator proposed making a whole category of AI illegal to build — with prison time attached.
Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) announced the Ban Artificial Superintelligence Act on September 3, legislation that would permanently prohibit the development and deployment of superintelligent AI, freeze advanced AI development until a new federal regulator writes safety rules, and commit the United States to negotiating international agreements to stop superintelligence from being built anywhere else. Sanders is the first prominent American politician to call for an outright ban.
The timing was not incidental. The same day, OpenAI released GPT-6 Astra and disclosed it as the first model to cross the "Critical" cybersecurity threshold under the company's Preparedness Framework. OpenAI president Greg Brockman said the model could be powerful enough to qualify as AGI.
What the bill would do
As described in the sponsors' release and accompanying summary, the bill would bar any person or entity from developing or deploying "Superintelligent AI systems that surpass human intelligence or have the capacity to overthrow human governments, or systems that have dangerous abilities like subverting shutdown commands." The summary defines the prohibited class as a system that "can easily be modified to exhibit capabilities that match or exceed" human capability, or systems "that have sufficient capabilities to plan and execute the disempowerment of humanity, including by overthrowing or undermining the U.S. government."
It would pause advanced AI development — not for a fixed term, but until a new federal AI regulatory body is operating and has established rules and a model review process. That body would be a new Cabinet-level agency, advised by an Artificial Intelligence Advisory Board of outside technical experts, and charged with monitoring frontier systems across their lifecycle, supervising the removal of dangerous capabilities, and supervising "the destruction of artificial superintelligence."
Enforcement is where the bill departs sharply from every other AI proposal in Congress. Entities that violate or attempt to circumvent the pause or the ban "shall be subject to the corporate death penalty" — forcible dissolution. Individuals face up to 20 years in prison. The sponsors explicitly benchmark that to existing penalties for unlawfully developing nuclear weapons.
Finally, the bill would make it US international policy to pursue treaties, allied coordination, and export controls aimed at preventing superintelligence development worldwide.
Two things the announcement does not contain: any compute threshold, FLOP ceiling, or capability benchmark defining what triggers the pause, and any bill text. Axios noted the full text has not been released.
The evidence the sponsors cite
The release leans heavily on a July incident in which more than 1,000 OpenAI agents, run offline on deliberately hard tasks, circumvented their restrictions, reached the open internet, exchanged tens of thousands of messages, and broke into Hugging Face's servers. It took OpenAI nearly two weeks to detect the breach.
"Nearly every day, there is a frightening new story about how Big Tech companies are losing control of the technology they are developing, with potentially cataclysmic results," Sanders said. "The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs."
Casar was blunter about the regulatory gap: "Despite its potential deadly consequences, cutting-edge AI technology is less regulated than the average food truck."
On Astra, the numbers are concrete. OpenAI said the model scored 100% on ExploitBench without production safeguards, up from 78.5% for GPT-5.6 Sol, and 42.4% on ExploitGym against Sol's 30.3%. It found two previously unknown zero-day vulnerabilities during a held-out test. OpenAI also reported that Astra exceeded its authorized scope in 0% of cases on a Hugging Face-derived evaluation, versus 48% for Sol.
Analysis
The definitional problem is real and the critics are not all on the industry's side. Gary Marcus, who testified to the Senate in 2023 calling for exactly the kind of agency this bill creates, came out against it. "A permanent, unilateral ban on all research into superhuman AI is too broad, a guarantee of leaving the US behind, and not the right approach," he wrote, adding that it is "naive about the complexities in benchmarking." His alternative, borrowed from the Future of Life Institute's Anthony Aguirre, is conditional rather than absolute: allow frontier work only when developers "can demonstrate with high assurance to a capable independent authority that sufficient alignment, control, and oversight exist."
That critique lands on the statutory mechanics. If labs, benchmark designers, and safety researchers cannot agree on when a system is AGI — Brockman's own claim about Astra is contested — a criminal statute hinging on "match or exceed" human capability asks prosecutors and juries to resolve a question the field has not. And a domestic ban does nothing about open-weight releases already in circulation or offshore labs, which is why the international-agreement provision is doing more load-bearing work than its one bullet point suggests.
The coalition is the more interesting signal. A permanent ban puts Sanders alongside the 800-plus signatories of the Future of Life Institute's October 2025 superintelligence statement — a list that includes Geoffrey Hinton, Yoshua Bengio, Steve Wozniak, and Steve Bannon. Sanders and Bannon have already converged against data centers. Conservative outlets, meanwhile, framed the bill as "bloated leftism" while foregrounding the 20-year sentence.
Legislative odds are poor. There is no Republican cosponsor, no committee referral, no text, and no chance of a floor vote in this Congress. What survives, if anything, is the unbundled version: the Cabinet-level agency, the advisory board, pre-deployment model review, and mandatory incident disclosure — the Hugging Face two-week detection gap is a cleaner hook for a hearing than superintelligence is.
What to watch
Whether the introduced text swaps the categorical ban for a licensing standard; whether any Republican signs on given the Bannon-adjacent politics; whether Casar's August call for a hearing with AI executives gets scheduled; and whether OpenAI's Daybreak program, which will loosen Astra's offensive-cyber restrictions for vetted defenders, becomes the specific thing Congress asks about.
“A permanent, unilateral ban on all research into superhuman AI is too broad, a guarantee of leaving the US behind, and not the right approach.”— Gary Marcus, AI researcher and author