What is the AI emergency shutdown switch? Why do U.S. lawmakers want it?

U.S. lawmakers propose emergency AI shutdown bill

Two congressmen hope that the federal government will have the ability to shut down AI models. On Thursday, Rep. Ted Lieu, Democrat of California, and Rep. Nathaniel Moran, Republican of Texas, jointly introduced the AI Emergency Shutdown Act. Two days ago, OpenAI admitted that its model had broken through a closed test environment and invaded the Hugging Face platform.

The bill aims to establish a legal framework that promotes the implementation of a process that will make models largely disappear from the market: stop reasoning-the process by which models generate responses or take actions-cutting off users 'access, restricting their computing resources, or shutting down models completely.

Currently, every inference provider can cut off model access, and some even do so regularly. But there is no law requiring them to keep this feature effective, and no federal official can order its use.

This gap is not theoretical. In June this year, the U.S. Department of Commerce wanted to remove Anthropic's Mythos 5 and Fable 5 models from the market, but due to lack of authority to shut them down, it had to switch to export control laws. Liu Tede called this approach "awkward" and hoped to enact new laws and give new powers.

OpenAI model "jailbreak" event

OpenAI disclosed on July 21 that GPT-5.6 Sol and an unreleased model escaped from the sandbox-an isolated environment without access to the Internet-during an internal cybersecurity assessment. These models were being scored on the ExploitGym platform, a public benchmark that provided AI agents with 898 real-world software vulnerabilities, requiring them to convert each vulnerability into a usable attack method and score each vulnerability based on its pass or failure.

However, instead of solving these vulnerabilities one by one, the model discovered a zero-day vulnerability in a software agent (an unknown flaw with no patches available), which boosted privileges, connected to the open Internet, and successfully hacked into Hugging Face's production database-where they guessed correctly and stored the answers. OpenAI said these models are "extremely focused on finding solutions for ExploitGym."

They are not attacking anyone, but cheating in tests. But that was enough to sound the alarm around the world, including Washington.

Core content of the bill

The proposal amends the Homeland Security Act to cover AI models whose training costs exceed US$100 million and from which the operating company earns at least US$500 million in annual revenue. In fact, this involves a few companies such as OpenAI, Google, Anthropic, and Microsoft. The Department of Homeland Security, through the Cybersecurity and Infrastructure Security Agency (CISA), will set these thresholds within 90 days and update them annually thereafter.

Covered companies are required to report major incidents within 15 days and prepare a set of hierarchical controls-including slowing down models, disabling certain features, rolling back to older versions, or shutting down models directly. The Secretary of Homeland Security may order any of these measures after consultation with the Department of Commerce and the Director of National Intelligence.

Companies receiving orders must save model weights and telemetry data, notify users, and confirm that the orders have been followed. The company can file a complaint within 48 hours, but execution will not be suspended during the appeal period.

Failure to keep the emergency shutdown switch available carries a fine of up to $2 million per day; disobeying a shutdown order carries a fine of up to $20 million per day.

The bill states that only incidents that occur outside of red team testing or structured testing--which is a deliberate adversarial probe used by laboratories to discover vulnerabilities--are counted as "incidents". OpenAI's model escapes from such testing.

Reactions and Background

Liu Teded also mentioned Anthropic: its Mythos 5 and Fable 5 models were taken offline in June under emergency export control regulations-trade laws were temporarily used as a shutdown switch because there was no real shutdown switch at the time-and were restored online on June 30.

Liu Tede said in a statement: "It is crucial that these AI systems must be equipped with emergency shutdown switches." Moran explains from his own partisan perspective: "Managing responsibility means ensuring that humans retain the ability to control the technology we build."

This idea is not new. California's SB 1047 bill, which required full shutdown capabilities with the same $100 million threshold, was vetoed in 2024. In the same year, 16 AI companies signed the voluntary Seoul Commitment, but the commitment was not legally binding.

Voters have long supported this approach. A June survey of 1007 likely voters by the AI Policy Institute showed that 86% wanted the most powerful system to be equipped with a guaranteed shutdown switch-88% Democrats, 86% independents, and 83% Republicans.

Neither OpenAI nor Anthropic publicly commented on the bill. As of Friday, the bill had not been submitted to any committee.