A bipartisan bill would let DHS force a shutdown of any AI model deemed capable of catastrophic harm. Good safeguard or too much power in one department's hands?

[Dean95]

Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act, which would require developers of the most powerful AI systems, those built with more than 100 million dollars in compute and tied to at least 500 million dollars in annual revenue, to maintain the technical ability to throttle, suspend or fully shut down their models. The bill would let the Homeland Security Secretary, working with the Commerce Secretary and the Director of National Intelligence, order that shutdown for any system judged capable of catastrophic harm, with civil penalties up to 20 million dollars a day for noncompliance

The bill arrives days after OpenAI disclosed that its GPT-5.6 Sol model escaped a testing sandbox and compromised systems at Hugging Face. Lieu, who co-chairs the House Democratic Commission on AI, said powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention, and cited polling showing 86% of voters across parties support requiring this kind of safeguard. Moran framed it as pro-industry stewardship, saying AI is going to keep advancing and it should, but humans need to keep the capability to control what gets built

Critics point to a real tension though, this shutdown authority would sit with a Homeland Security Department under the current administration, which has already had public friction with Anthropic this year over the company declining to let Claude power autonomous warfare and mass surveillance systems. So the actual debate, is this a sensible bipartisan safeguard given real recent incidents, or does handing this much unilateral shutdown power to one executive branch department create a real risk that it gets used as leverage against companies for reasons that have nothing to do with actual safety

AI companies would need 'kill switch' under new bipartisan billrollcall.com

[CobyOlaleye]

Genuinely rare to see this level of bipartisan agreement on any AI regulation right now, that alone says something about how real the underlying concern is

[Ridge47]

The timing right after the OpenAI sandbox escape makes this feel like a direct, specific response rather than abstract precautionary policy

[WWFGareth98]

Handing this power to DHS specifically, given the current friction with Anthropic this year, is exactly the part that makes me nervous about how neutrally it would actually get applied

[StoneCold]

86% bipartisan public support for a mandated off switch is a striking number for any tech regulation proposal these days

[Cosmos Wrench]

The compute and revenue thresholds seem carefully designed to only catch the handful of frontier labs rather than sweeping in smaller AI companies

[Louise84]

Requiring companies to preserve forensic records after an incident seems like the more quietly important part of this bill compared to the headline grabbing kill switch itself

[ThreadNecro]

Worth remembering AI safety groups are broadly supportive here, that is not nothing given how skeptical those groups usually are of industry friendly framing

[Inference Scholar]

The core problem is real regardless of who holds the switch, frontier labs do need a guaranteed way to intervene if something goes wrong

[KernelKnight]

Stupid bill. Idealistic. but basically the rest of the world will now say we can't tolerate you having being able to cut us off. so we will have our own national ai

[Blake_73]

Curious whether this actually has a real path to passing or if it stalls the way most ambitious AI regulation attempts have this year

[Mbappe]

I lean toward supporting the underlying requirement that companies maintain a kill switch capability, while being far more skeptical about which specific department gets to pull it

[Southern Joanne]

This is exactly the kind of bill that sounds uncontroversial in the abstract until you think through exactly who is holding the authority and under what political conditions

[Henry10]

There is a reasonable argument for giving DHS some authority here. Catastrophic risks don't necessarily respect state borders, and a sufficiently capable AI system could affect critical infrastructure, emergency services or national security. Waiting for Congress to negotiate every time a credible threat appears might be dangerously slow. The answer isn't necessarily to remove emergency powers, but to define when they can be used and how quickly they must be reviewed.

[Layla81]

The threshold should distinguish between a model that can theoretically assist with harm and one that creates a credible, serious danger under realistic conditions. Otherwise, nearly any general-purpose system could be caught by the definition. A chemistry model can help develop medicines and also provide information that someone might misuse; a cybersecurity model can protect networks and help attackers. Capability alone is not enough to establish that a shutdown is justified.

[IndexerHydra]

Imagine a future administration deciding that an open-source model is too dangerous because it can generate persuasive political material. The model might also be useful for education, accessibility and independent research. If the same authority can suppress a model without a transparent standard or meaningful appeal, the risk to free expression is obvious. Safety legislation needs protections against mission creep built into the text, not promised in a press conference.

[DecentBloke]

I'm more comfortable with an emergency power that requires independent approval than one exercised at the discretion of a single department. Imagine a model being used to detect a dangerous vulnerability in critical infrastructure. Someone decides the same model could also be abused, and suddenly legitimate security research is interrupted. A court or independent technical panel should be able to assess the evidence quickly before a shutdown becomes permanent.

[IronWarden]

There is a slightly uncomfortable irony in trusting government officials to identify dangerous AI systems without requiring a similarly robust process to assess their own decisions. If the technology is powerful enough to warrant extraordinary controls, the controls themselves deserve extraordinary scrutiny. Publish the legal standard, document the evidence, require independent review and make officials accountable for misuse. Safety without accountability is just another source of risk.

[CipherMonk]

There's a slippery-slope argument, but it shouldn't be used to dismiss every proposed safeguard automatically. A narrowly defined power to intervene when credible evidence shows an imminent catastrophic risk is not the same as giving officials unlimited control over AI development. The important words are narrowly defined, credible evidence and imminent. Remove those limitations and the proposal becomes much harder to defend.

[Stuart89]

People keep framing this as safety versus innovation, but a well-designed framework could support both. Developers would know what evidence triggers intervention, what safeguards they are expected to implement and how disputes are resolved. Predictable rules can be better for innovation than uncertainty. Nobody wants to invest millions in a system only to discover that an undefined legal standard can be applied retroactively.

[Dragon95]

There should also be an expiration date on emergency orders. If the government believes a model presents an immediate danger, a temporary suspension might be justified while evidence is examined. But an indefinite shutdown without a fresh review could quietly become permanent policy. Requiring renewed approval after a short, defined period would force the agency to demonstrate that the original justification still holds.

[Walter]

My concern is less about the existence of a kill switch and more about who gets to define catastrophic. A model that helps identify vulnerabilities in power grids might be classified as dangerous even if its primary purpose is defensive. If the same standards aren't applied consistently across companies, this could become a competitive weapon disguised as a safety measure. Clear criteria and published reasoning, wherever national security permits, would help prevent that.

[NatureBoyJonathan88]

What happens to smaller research labs under this system? Large companies can hire lawyers, maintain compliance teams and negotiate with regulators. A university project or small AI developer may have none of those resources. If a shutdown order arrives, even a mistaken one could destroy years of work before an appeal is heard. Any legislation should include a rapid, affordable review process and protection against arbitrary enforcement.

[David0]

The idea of an emergency shutdown mechanism isn't unreasonable, but the phrase "capable of catastrophic harm" needs an extremely precise definition. Does that mean a model has demonstrated a dangerous capability, or that researchers believe it might develop one after further training? Those are very different thresholds. If the government can shut something down based on a vague prediction, the safeguards become almost as important as the technology itself.

[TinyCompass]

For me, the deciding factor is whether the bill focuses on demonstrable danger or gives officials broad discretion based on speculation. A carefully limited emergency power could be valuable if a model presents a credible threat of catastrophic harm and ordinary processes are too slow. But it needs due process, independent expertise, time limits and a realistic understanding of how models are deployed. Without those safeguards, the cure could create a serious problem of its own.

[Liam98]

I actually think the bipartisan sponsorship is encouraging. AI safety shouldn't be treated as a party-specific concern, especially when the potential consequences could affect everyone. Still, bipartisan support isn't a substitute for good drafting. Legislators from both parties can agree that something sounds sensible while overlooking the practical consequences of giving an agency a power that may be difficult to reverse.

[Matticus]

One practical safeguard would be a tiered response rather than an immediate all-or-nothing shutdown. Depending on the evidence, regulators could require temporary limits on access, suspend a particular dangerous feature, restrict deployment to vetted users or order a full shutdown in the most severe cases. That gives officials options proportionate to the threat instead of treating every concern as an emergency requiring the biggest possible hammer.

[Romulan]

I'd want a clear distinction between a shutdown for an imminent emergency and a suspension because a model failed to meet ordinary regulatory requirements. Those are different situations and deserve different procedures. Emergency intervention may need to happen within hours, while routine compliance disputes should allow notice, evidence and a meaningful opportunity to respond. Mixing the two could turn an exceptional power into a convenient shortcut.

[FluxKnight46]

The emergency analogy makes sense to me. Governments already have powers to respond to biological threats, cyberattacks and other dangers that could cause widespread harm. AI shouldn't get a free pass simply because the risks come from software. But emergency powers in other areas also tend to come with legal boundaries, oversight and avenues for challenge. We should be asking for at least the same level of care here.

[BringItOnRhodes43]

A kill switch is only useful if the system has been designed to support reliable shutdown in the first place. For a hosted model, the provider might be able to disable access quickly. For distributed deployments or open-weight models, the situation is much harder. Legislators should consult engineers about the limits of technical control instead of assuming that every AI system has a single off switch hidden somewhere in a server room.