House Proposal Requires AI Shutdown Controls
Reps. Ted Lieu, D-Calif., and Nathaniel Moran, R-Texas, are preparing to introduce the AI Kill Switch Act, a bipartisan House proposal that would authorize the Department of Homeland Security to order certain AI systems shut down or throttled. Politico reports that covered companies would need incident-reporting processes and technical controls for suspending models or user access, with penalties reaching $20 million per day.
Reps. Ted Lieu, D-Calif., and Nathaniel Moran, R-Texas, are preparing to introduce the bipartisan AI Kill Switch Act, which would authorize the Department of Homeland Security to order some AI companies to shut down or throttle powerful models. According to Politico, the proposal would require covered firms to report incidents and maintain the technical capability to shut down, slow, or suspend access to their systems.
The proposed legislation is not law. It is one of several congressional efforts focused on risks from advanced AI systems, and Politico reported that it was scheduled for introduction on July 23.
Thresholds and emergency authority
The Verge reports that DHS could issue an order after consulting the secretary of commerce and the director of national intelligence. The authority would apply in specified "loss-of-control" scenarios, including incidents involving at least 10 deaths, more than $100 million in economic damage, or a model attempting to conceal shutdown controls.
Politico reports that the bill would generally cover models developed with at least $100 million in computing power and companies generating at least $500 million in annual revenue from the technology. Violations of emergency shutdown orders could carry civil penalties of up to $20 million per day, according to both Politico and The Verge.
The proposal follows OpenAI's disclosure of what Politico described as an autonomous hack during an internal evaluation, involving two advanced models and Hugging Face. Politico reported that the incident intensified lawmakers' concern about the cyber capabilities of frontier models.
What the mandate would operationalize
If enacted, the measure would turn an often-discussed AI safety concept, operator control over deployed systems, into a compliance obligation for a defined set of high-cost models. The Verge reports that the required controls would include the ability to turn off systems, reduce their operating capability, and suspend user access.
For ML and platform teams, comparable regulatory regimes commonly create work across several technical layers:
- •Model and service inventory, including identifying which deployed versions fall within a statutory threshold.
- •Privileged operational controls that can disable inference, rate-limit capacity, or revoke customer access.
- •Incident telemetry and escalation procedures capable of supporting mandatory reporting.
- •Auditable access controls, so emergency actions can be executed and later reviewed.
Those are general implementation patterns rather than evidence of any particular company's existing capabilities. The bill's practical reach, if it advances, would depend on how lawmakers define compute, model scope, incident reporting, and the evidentiary standard for a DHS emergency order.
Key Points
- 1Proposed DHS authority would combine emergency intervention with mandatory shutdown controls for certain high-cost frontier models.
- 2The bill sets reported compute and revenue thresholds, concentrating obligations on companies developing and commercializing the largest systems.
- 3Comparable compliance programs require auditable model inventories, privileged service controls, incident telemetry, and documented emergency-response procedures.
Scoring Rationale
The proposal could establish consequential operational and reporting requirements for developers of the largest AI models if enacted. It is proposed legislation rather than an active rule, but its emergency shutdown provisions are directly relevant to frontier-model deployment and governance teams.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
