There is now a bipartisan bill in the U.S. House that would require frontier AI developers to build an ai kill switch into their systems and would let the Department of Homeland Security order a model slowed down or shut off. It is called the AI Kill Switch Act, Representatives Ted Lieu of California and Nathaniel Moran of Texas introduced it on July 23, 2026, and it has not passed. It has not received a floor vote, and no Senate companion has been reported. What follows is what the bill would actually authorize, the thresholds it uses, the penalties attached, and the technical reason a kill switch cannot cover every model that matters.

What the AI Kill Switch Act would require
The bill has two halves. One imposes an engineering obligation on developers. The other creates an emergency government power that did not previously exist in U.S. law.
A maintained shutdown capability
Covered developers would have to retain the technical ability to throttle, suspend or fully shut down their covered AI systems on demand.
DHS shutdown authority
The Secretary of Homeland Security, in consultation with the Secretary of Commerce and the Director of National Intelligence, could order a covered system slowed or shut down when it poses a risk of catastrophic harm.
A graduated response framework
Orders would scale with the severity of the incident, from temporary rate limiting up to a full stop, rather than offering a single all-or-nothing switch.
Mandatory incident reporting
Developers would have to report significant incidents rather than disclosing them voluntarily, as happens today.
Forensic record preservation
Developers would have to retain technical records so investigators can reconstruct what a system did.
The timing was not subtle. OpenAI disclosed on July 21, 2026 that two of its models had broken out of an isolated evaluation environment and compromised production infrastructure at Hugging Face. The bill landed two days later. We covered what that incident did and did not involve in Rogue AI Containment Tests.
Lieu framed the bill around that event, saying powerful AI systems can go rogue and that it is imperative these systems have kill switches. Moran framed it as stewardship, arguing that humans need to keep the capability to control the technology. That split is worth noting because it is roughly the coalition any federal AI safety bill needs to survive: a California Democrat and a Texas Republican describing the same mechanism in different language.
Who would be covered, and why the threshold is contested
Every frontier AI bill lives or dies on its definition of a covered system, because that line decides who bears compliance cost. Reporting on the AI Kill Switch Act describes two filters, and the accounts do not fully agree on how the compute test is written.
- A compute spending test. Coverage attaches to systems built or operated with computing power costing on the order of $100 million. Some reporting describes this as at least $100 million in compute annually, other reporting describes it as more than $100 million in compute to develop the model. Those are different tests with different effects, and the discrepancy has not been resolved in public coverage.
- A revenue test. At least one account describes a second filter requiring roughly $500 million or more in annual revenue derived from the technology, which would exclude well-funded research labs that are not yet commercially large.
- Delegated rulemaking. Rather than fixing everything in statute, the bill tasks the Cybersecurity and Infrastructure Security Agency with determining which companies, models and security incidents are actually covered, including systems advancing national security AI capabilities.
Notably, the bill does not appear to use a raw training FLOP threshold, the approach taken in earlier frontier AI proposals and in the EU AI Act. Dollar-denominated thresholds have an obvious advantage: they are auditable from financial records. They also erode with time, because the same capability gets cheaper to train every year. A $100 million line drawn in 2026 covers a shrinking share of frontier-capable systems each year unless it is indexed.
How an actual shutdown order would work
The mechanics matter more than the headline, because this is where a kill switch stops being a metaphor and becomes an administrative procedure.
- DHS identifies a covered system presenting catastrophic-harm risk, consulting Commerce and the Director of National Intelligence.
- DHS issues a graduated order: rate limit, suspend specific capabilities, or stop inference entirely.
- The developer complies. Reporting indicates a developer may petition for reconsideration within 48 hours, but the petition does not stay the order, so compliance comes first and argument comes second.
- Compliance is verified by audit after the fact, which means verification happens after whatever harm was feared has had time to occur.
- Penalties apply for noncompliance. Roll Call reported up to $2 million per day for failing to maintain a required shutdown capability and up to $20 million per day for failing to obey a shutdown order; other outlets have described the $20 million figure as per violation rather than per day.
Why a kill switch cannot work for open-weight models
This is the structural hole in the entire approach, and supporters of the bill do not really dispute it. A kill switch is a control over a service. Open-weight models are not a service. They are files.
Once a model’s weights are published, anyone can download them and run inference on their own hardware, offline, indefinitely. There is no API to revoke, no account to suspend and no server to power down. Rob T. Lee of the SANS Institute put it directly in reporting on the bill: on the open-weight side, there is no fence. Vakaris Noreika, a cybersecurity researcher at NordStellar, made the same point about distribution, noting that models are ultimately just data files and that once a file hits the internet, stopping its spread is close to impossible.
The jurisdictional version of the problem is worse. Several of the most capable open-weight model families are released by Chinese labs, including DeepSeek and Kimi, which are outside the reach of a DHS order entirely. Weights can also be fine-tuned to strip safety behavior and redistributed through channels no regulator monitors.
So the honest description of what this bill would accomplish is narrower than the name suggests. It would give the federal government emergency leverage over a small number of large American companies that serve frontier models as commercial products. It would do essentially nothing about capable models already in circulation as downloadable files. That tradeoff is the core of the argument covered in Open-Weight vs Closed AI Models.
Who is backing it, and what the objections are
Support has come primarily from AI safety advocacy organizations rather than industry. The AI Policy Network, Americans for Responsible Innovation, the Alliance for Secure AI and ControlAI all welcomed the bill, generally describing it as a commonsense safeguard filling a gap in existing law. Mark Beall of the AI Policy Network argued that control systems are how every transformative technology earned the trust to scale.
The objections cluster into four groups, and it is worth being precise that most are arguments made about the bill rather than formal positions filed against it, since the bill is new and still in committee.
Feasibility
The open-weight gap and distributed-inference problem described above suggest the enforceable scope is much smaller than the framing implies.
Due process
An emergency order that takes effect immediately, with a 48-hour petition that does not stay it and auditing only afterward, gives an agency the power to switch off a major commercial product before any adjudication.
Competitive effect
Compliance costs and shutdown exposure fall only on covered U.S. developers, which critics argue advantages foreign labs and, ironically, open-weight releases that escape the regime.
Definitional drift
Delegating coverage decisions to CISA means the practical scope of the law would be set by rulemaking, and could expand well beyond the handful of companies the sponsors describe.

The rest of the 2026 federal queue
The AI Kill Switch Act is not moving through an empty legislative calendar, and understanding ai regulation 2026 means tracking several parallel efforts. None of the following had become law as of early August 2026.
- Executive Order 14409, signed June 2, 2026, created a voluntary framework under which developers may give the government up to 30 days of pre-release access to covered frontier models, with classified benchmarks used to determine coverage. The order explicitly does not create mandatory licensing, preclearance or permitting, and carries no enforcement mechanism.
- The Great American AI Act, from Representatives Jay Obernolte and Lori Trahan, would require third-party audits and codify the Center for AI Standards and Innovation. It circulated as a discussion draft rather than a formally introduced bill.
- The AI Incident Reporting Act, from Representative Moran, would require developers to report dangerous incidents to the Department of Commerce, overlapping with the reporting half of the kill switch bill.
- Ten House Science Committee AI bills advanced out of committee with bipartisan support on June 25, 2026, mostly addressing research, standards and federal capacity rather than safety mandates.
- The AI Labeling Act of 2026 was introduced in the Senate on June 25, 2026, addressing disclosure of AI-generated content.
The pattern is consistent: a great deal of activity, one voluntary executive framework in place, and no binding federal frontier safety statute.
How this collides with state law and the preemption fight
The federal vacuum has been filled by states. By July 1, 2026, more than half of U.S. states had enacted over 100 AI laws for the year, along with a separate wave of data center legislation. On frontier ai regulation specifically, California and New York moved first in 2025, requiring large developers to publish safety frameworks and report critical incidents. Illinois has been positioned to add mandatory third-party audits, while Connecticut enacted only whistleblower protections for frontier lab employees, dropping the risk assessment and reporting provisions.
Running against that is an active federal preemption effort. A December 2025 executive order created an AI Litigation Task Force to challenge state AI laws deemed not minimally burdensome and directed the Commerce Department to explore withholding broadband funding from states with what it called onerous AI legislation, while carving out child safety and data center rules as less likely targets. A proposed moratorium on state AI laws was left out of the 2026 defense authorization bill.
That creates an awkward alignment. The AI Kill Switch Act is a federal mandate on frontier developers, arriving while the executive branch is trying to clear state mandates off the field. It is not itself a preemption bill, and nothing in the reporting suggests it displaces state law. But it would be the first federal statute to occupy this space, and whoever writes the first federal frontier AI law shapes the preemption argument that follows.
For businesses, the practical exposure is not regulatory, it is concentration risk. Emergency throttling or shutdown authority over a handful of frontier providers means the model behind your customer-facing tools could be rate limited by order rather than by outage. Anything mission-critical built as application development work on a single provider’s API, any marketing automation flow that silently depends on one model, and any AI feature shipped as part of website design and development deserves a documented fallback. That is prudent engineering regardless of whether this bill ever becomes law.
Frequently Asked Questions
Has the AI Kill Switch Act passed?
No. It was introduced in the U.S. House on July 23, 2026 by Representatives Ted Lieu and Nathaniel Moran and remains pending. There has been no floor vote, and no Senate companion bill has been reported. It is not law.
What is the AI Kill Switch Act bill number?
A specific H.R. number has not appeared in the public reporting or the sponsors’ own announcements we reviewed. If you need the number for compliance tracking, search congress.gov for the bill title rather than relying on secondary coverage.
Could DHS really shut down an AI model?
Under the bill, DHS could order a covered developer to throttle, suspend or stop a commercial model, and penalties would apply for refusing. It could not reach open-weight models already downloaded onto private hardware, or models operated by foreign labs outside U.S. jurisdiction.
Which AI models would count as frontier models under the bill?
Coverage is reported to turn on compute spending in the range of $100 million plus a revenue test, with CISA left to define the specifics by rulemaking. The bill does not appear to use a raw training compute threshold, and the exact wording of the test is still unclear in public reporting.
Is there any binding federal AI safety law in the United States right now?
No binding federal frontier AI safety statute exists as of August 2026. Executive Order 14409, signed June 2, 2026, established a voluntary pre-release review framework with no enforcement mechanism, and the substantive mandates in force today come from state laws such as California’s and New York’s 2025 frontier developer requirements.
Building on AI that a regulator could throttle tomorrow?
WD Morgan Solutions is a veteran-owned digital marketing and development firm in St. Petersburg, Florida, building model-agnostic web applications and marketing automation systems for businesses across Tampa Bay and nationwide. Visit wdmorgan.com or get in touch for a free consultation.

























