AI Kill Switch Act: US Government Could Order AI Shutdowns

A newly proposed US bill called the AI Kill Switch Act would give the Department of Homeland Security the authority to order AI companies to block user access, restrict specific capabilities, or shut a system down entirely when that system is judged capable of causing catastrophic harm. The bill would require AI developers to build the technical infrastructure for such shutdowns before they are needed. Non-compliance would carry fines of up to $20 million per day.

What happened

Detail	Fact
Bill name	AI Kill Switch Act
Authority granted to	Secretary of the Department of Homeland Security
Law amended	Homeland Security Act of 2002
Fine for non-compliance	Up to $20 million per day of violation
Actions the government could order	Block user access, disable or restrict a capability, or full system shutdown

The bill’s sponsors announced the proposal and stated that it targets AI systems “that can cause catastrophic harm.” If passed by Congress, it would apply to the current Trump administration and all future administrations, making it a permanent legal framework rather than a temporary executive measure.

Crucially, the bill does not just grant shutdown authority after the fact. It would require AI developers to proactively build technical capabilities, such as throttling mechanisms and full kill switches, into their systems so the government can act quickly if ordered to do so.

Why it matters

This bill represents one of the most direct attempts yet to impose hard operational controls on AI companies at the federal level. Most AI regulation discussions so far have focused on disclosure, risk assessment, or liability. This proposal goes further by mandating government-accessible off switches built into the technology itself.

The $20 million per-day fine structure is significant. For context, that figure escalates fast: a company refusing to comply for a single month could face over $600 million in fines. That kind of penalty creates strong financial pressure to comply, even for large AI labs.

The choice of DHS as the authority is also notable. Placing this power under the Secretary of Homeland Security frames advanced AI as a national security issue rather than a consumer protection or competition issue. That framing tends to result in broader and faster-moving enforcement powers.

For businesses that rely on AI tools, including platforms built on models from OpenAI, Anthropic, Google, or others, a government-ordered shutdown of an underlying AI system could disrupt operations with little or no warning. If you have built workflows or customer-facing products on top of third-party AI APIs, this is the kind of regulatory risk worth factoring into your architecture decisions now. Our work on AI integration for businesses increasingly includes questions about redundancy and fallback options for exactly this reason.

Is this bill likely to pass?

That is genuinely unclear at this stage. The bill has been introduced but has not yet been approved by Congress. Similar AI safety proposals in the US have stalled before. However, the current political climate has made AI risk a bipartisan talking point, which gives this bill slightly more room than previous attempts.

The requirement for pre-built kill switches will likely face pushback from AI developers, who may argue that mandating such infrastructure creates its own security risks or competitive disadvantages. Expect lobbying from major AI labs as the bill advances.

You can follow ongoing coverage of AI policy developments on the Lumien news feed as this legislation moves through Congress.

Our take

The intent here is understandable. If a genuinely dangerous AI system were causing harm at scale, there should be a mechanism to stop it. The problem is that “catastrophic harm” is a phrase that can be defined narrowly or very broadly depending on who is in office and what pressure they are under.

Handing a single cabinet official the authority to shut down an AI system, with no mention of judicial review or independent oversight in what has been reported so far, is a design worth scrutinising. The fines are large enough to compel compliance before any court challenge could run its course. That asymmetry matters.

For business operators, the practical concern is not abstract. If your product or internal tooling depends on an AI API and that API gets ordered offline, you may have hours or days to respond. Building with provider diversity and local fallback options is no longer just a good engineering practice. It is starting to look like basic operational resilience.

What to do about it

1. Audit which of your business processes depend on a single external AI provider’s API.
2. Identify which workflows would break immediately if that provider went offline for 48 to 72 hours.
3. Evaluate whether a secondary provider or a locally hosted model could serve as a fallback for your most critical use cases.
4. Watch the bill’s progress through Congress and revisit your risk assessment if it advances to a committee vote.

Regulatory risk is now part of the AI stack. Plan for it the same way you plan for server downtime.

Source: Ars Technica · AI