Video
In this episode of Tech News this Week, host Kelsey Sung is joined by Richard Livingston to unpack the recent discourse around AI "kill switches," including the feasibility of an AI kill switch and the emerging legislation calling for such shutdown capability and increased AI oversight.
Be sure to follow Tech News this Week on Podbean, Spotify, Amazon or wherever you get your podcasts.
Amid the wider debate of whether AI development needs to slow down for safety's sake, another question has emerged: Is it time for an AI kill switch?
Several recent pieces of legislation have proposed the concept of a mandatory kill switch for AI systems. In July 2026, lawmakers introduced the bipartisan H.R. 9917 -- or the AI Kill Switch Act -- which would require AI agent developers to have shutdown capability and a mandatory incident reporting process.
In September, Sen. John Kennedy introduced S. 5417 -- or the AI Emergency Button Act -- which would similarly require AI developers to install an emergency kill switch. In the same week, in response to recent AI security incidents, California Gov. Gavin Newsom issued an executive order that advances the creation of an 'AI kill switch' and establishes independent oversight of AI developers.
But how realistic is the idea of an AI kill switch?
"[Experts] say that we can't have an AI kill switch for the same reason that the internet doesn't have an off button. The ecosystem is just too complex," Richard Livingston, editor at TechTarget Cybersecurity, said on a recent episode of Tech News this Week. "There's no one place to go ahead and knock this thing out."
A complex ecosystem, Livingston noted, complicates the downstream impact of triggering such a kill switch.
"I think there's going to be a blast radius," he said. "If you have to shut down AI models that systems are dependent on downstream of you, your vendors could be affected. Other systems that you certainly never expected are going to be affected. Somebody's going to get hurt in that."
Watch this episode of Tech News this Week to dive deeper into what an AI kill switch could realistically look like, the arguments for and against such a feature, and the cybersecurity reckoning CISOs face in the wake of this debate.
Kate Murray is an editor for TechTarget's Reader Engagement and Outreach team. She joined the company as an associate managing editor of e-products in 2020 and has also served as managing editor for TechTarget's Infrastructure sites.
Transcript - AI kill switch: Can we really shut down artificial intelligence?
Kelsey Sung: Can AI be stopped? That is the crucial question as lawmakers debate a mandatory AI kill switch.
Richard Livingston, site editor on TechTarget Cybersecurity, has been diving into that question and more and is here to share what he's found so far about the AI kill switch.
Richard, thanks for coming onto the show today.
Richard Livingston: No, thanks so much for having me, Kelsey.
Sung: Our pleasure. So, the phrase 'AI kill switch' has kind of become almost colloquial at this point in the last couple of weeks, but what does it actually mean? Like, when lawmakers talk about instituting an 'AI kill switch,' what is that?
Livingston: Yeah, so it's really a broad security concept and it's intended to restrict or to halt an autonomous AI model. And this is really not one thing. It could be a whole set of actions. It could range from really cutting the power to just a gradual step down that would restrict the capabilities of an AI model. It could isolate it from a network. It could revert it to a safe state. But basically what it is, it's a little bit of a breather that would let analysts stop the actions of the agent and step back if it starts acting unpredictably and really investigate what's going on there without it continuing to, you know, do things like we've seen in the Hugging Face incident.
Sung: Yeah, OK. So, it's not like a big red button, which is what I was imagining. That makes more sense.
Livingston: Yeah, absolutely.
Sung: So, there is this raging debate over whether stopping an AI completely is even possible. We're talking specifically about AI agents here, right? Specifically, autonomous AI agents, kind of like the OpenAI that -- not kind of, like the OpenAI model that attacked Hugging Face, as you just mentioned.
Livingston: Yeah, but so far, and these behaviors that we've seen so far are demonstrated by really the advanced frontier models right now. But the danger, of course, is that these are going to become more common and really more of the standard for what AI agents are doing.
Sung: Yeah, OK. So, can you just run us through the players in this debate, the arguments for and against legislation like this, and what does the landscape look like right now in this kind of conversation?
Livingston: Yeah, sure thing. So, now there is raging debate in Washington over this as far as really where we are with that discussion. We have a bipartisan House bill sitting out there right now. It's H.R. 9917, the AI Kill Switch Act, and this would require developers of AI agents to have shutdown capability, to have mandatory incident reporting if these agents start acting unpredictably. And then, in my opinion, the most interesting one is also this would have a possible government override that, if this agent is acting in a way that we don't like and the developers are not shutting down, Uncle Sam will come and shut it down for you until we can figure out exactly what's going on.
There are other things. This is sitting in the House right now, but the Senate -- just in this last week -- Senator John Kennedy proposed a similar bill in the Senate. And then, on the other side of the country, in California, they're really out ahead of a lot of these things, which I guess would make a lot of sense, considering that's where Silicon Valley is. But Governor Newsom, he has issued an executive order that would advance the creation of AI kill switches -- of these algorithms that would go ahead and halt your agents -- and then also have mandatory, human, independent oversight of AI. And this is on top of some things that they've already done. They already have a mandatory AI safety assessment for new models coming out, and they have a state registry of AI auditors.
So, a lot going on. And then, also, I should mention that on the other side of the pond, in the EU, they have Article 14 right now. And that probably will not be in full effect until 2028, but it is the same -- it is basically their version of the Kill Switch Act.
So, a lot of players, a lot of things going on right now. But, you know, you asked also about the argument against. And the experts with whom I speak, basically, they say that we can't have an AI kill switch for the same reason that the internet doesn't have an off button. The ecosystem is just too complex, you know? You have AI distributed across models, the cloud, vendors, apps, you know, they're spawning subagents. You know, really, what you would be looking at feasibly is kind of like Whac-A-Mole. There's no one place to go ahead and knock this thing out.
And then, you know, the one other thing I'll say about this is, also this is an untested concept. There has never been a live-fire exercise with an AI kill switch. Nobody has used one yet, so will it work? I don't know.
Sung: Yeah, we kind of don't know until we're in the situation where we might have to rely on one, which is a scary thought.
So, you kind of touched on this, but there's security leaders who are saying, as mentioned in your article on the subject matter, that this is kind of a dangerous illusion that we can even have a kill switch. But also, you mentioned that 86% of voters across all parties want guaranteed off switches for AI systems. So, it feels like there's a bit of a disconnect between what the public understands about AI kill switches and what the experts are saying is possible.
Can you talk a little bit more about that gap and what the technical reality actually looks like? You've touched on that a little bit -- but, like, why is there such a disconnect between reality and what the public is expecting from these kill switches?
Livingston: Yeah, and that's a really interesting question. And I think -- so many things. I think this is partly driven by fear. I mean, I think advanced AI lives rent-free in our heads. You know, if I say, "Would you like to play a game?" [or] "Skynet is active," you know exactly what I'm talking about. We have 50 years of sci-fi that have been preparing us for this moment.
And then just a few weeks ago, Jacob Coxon, in his very public and viral resignation from Anthropic, he said -- with, I'll say, some degree of hyperbole -- that, you know, AI has a good chance of killing us all by the end of the decade. There is, you know, there's just a lot of headline-grabbing fearmongering going out there, and the public outcry is reflecting that. And our government, as they tend to do, I think they kind of figure, well, something rather than nothing is the way to go here.
And, you know, I think really, you know, in terms of the realism of something like this working? Yes, a kill switch is feasible. It could go ahead and prevent an agent from acting in unexpected and potentially harmful ways. The problem is, because of those dependencies that I talked about, when you hit that kill switch, you're probably going to break something else. There's going to be a big downstream impact there. So, the experts are saying we should incorporate it in our models, but really as a component of last resort. Before that, you should have all your common-sense cybersecurity governance, where you are restricting access of AI to other systems the same way you would any other potential threat agent.
Sung: Yeah, so creating more of a -- to use the kind of buzzword now -- an ecosystem of cybersecurity instead of relying on this kill switch to solve everything as a silver bullet.
Livingston: Absolutely, one element of a much broader plan.
Sung: Yeah. And hopefully one that never needs to be used anyway.
Livingston: Exactly.
Sung: Yeah, so, I mean, kind of along those lines, you already mentioned Governor Newsom in California. This is not just an abstract kind of concept anymore -- this legislation around AI kill switches. Governor Newsom in California has already instituted an executive order along those lines.
What has the response to that executive order been? I'm curious, especially because that might tell us something about how this conversation might go as it becomes more concrete. So, like, what are cybersecurity experts that you've spoken to saying about that specific executive order, the pros and cons of it, and how has it -- if at all -- influenced the national conversation?
Livingston: Yeah, so there's, as I mentioned, you know, they see some practical fallacies in the argument for a kill switch. But I think where they see the benefit of this debate is, basically, this is forcing a reckoning of the cybersecurity industry to map their dependencies; to go ahead and look at the idea of layered controls of resilience, of that good cybersecurity governance, and all these things that we've been building over the last 25 years in terms of cybersecurity, zero trust, restricted permissions, clear incident response, making the assumption that your employees could be the threat, therefore there's a mechanism to, you know, restrict their permissions.
If you're going to have an autonomous agent in your systems, it's going to be helping you with a lot of things. But the same way a human being could start doing things that you certainly don't expect them to do, an AI agent could do the same thing. So, just make sure that you've doubled down on, you know, on that good governance. And I think that's forcing CISOs to really look at their systems and, you know, prepare for a worst case, which I think is a healthy exercise.
Sung: Yeah, yes, just in general, a healthy approach. So, I'm curious. As you look to the kind of -- I'm not even, I'm just going to say the next couple of months. We can't say, you know, years down the line at this point. For the next couple of months, as a reporter, what are you going to be focusing on? What questions are you hoping to find answered, if any, that can be answered in a couple of months, and just what are you going to be keeping an eye on as this evolves?
Livingston: Yeah, so honestly, I think legislation around kill switches will absolutely go into effect. They're going to build them. They're going to be in those systems and perhaps not in the next couple months, but, you know, we've been surprised at the pace of this thing before. At some point, somebody's going to have to enact it. And what I'm interested in is what happens then. And I think it's going to be one of two things.
The first is that it doesn't work, that you have these advanced agents that have been able to circumvent it, that they've spawned subagents, that there's just too much downstream, and you have not totally been able to contain that model.
The other side of that is: I think it could work, but I think there's going to be a blast radius. If you have to shut down AI models that systems are dependent on downstream of you, your vendors could be affected. Other systems that you certainly never expected are going to be affected. Somebody's going to get hurt in that. So really, I think we're -- at some point, we're going to understand what happens when you hit that kill switch. And then, when we have those answers, I think that will spawn a whole new debate about, you know, what do we do now?
But, you know, they say, 'May you live in interesting times,' and we certainly do.
Sung: Yeah, we won that.
Livingston: Yeah, yeah.
Sung: I think something that's going through my head, given the news yesterday about Trump and the AI leaders kind of signing this gentleman's agreement, it seems like, for self-regulation. Does that factor at all into this conversation about AI kill switches and who has the power to regulate AI? And if this AI kill switch goes into effect, if that kind of legislation becomes reality, how does that coincide with this narrative around self-regulation amongst these AI companies?
Livingston: Yeah, so I, you know, I think really what we're probably looking at is a government-industry partnership on this. I mean, those exist, you know, in transportation, those exist in so many other industries. And I think the best possible outcome is if we can have the industry self-regulate. Now, you know, nobody's going to grade their own homework. You can't just give Anthropic the decision as far as whether they're, you know, they're going to go ahead and contain themselves.
But if you have a group of frontier AI developers, a peer industry group who are meeting, who are reviewing each other's work -- you know, within reason, of course, there's competitive arguments here. But I think that the industry can self-regulate here, in which case, hopefully, the government wouldn't need to get as involved. But I think that should absolutely, you know, be a component of it.
And then, you know, and then there are all sorts of broader questions too. If, you know, if we're doing it here, what's China doing? What's Russia doing? There's a lot of players out there. But, back to your original question, I think that this will be a partnership. But my hope is that the industry can find a way to -- I mean, they're already speaking up and saying, you know, "Somebody stop me." You know? They realize what's happening here, and they're asking for government oversight.
Sung: Yeah. Man, it is a lot of layers to this conversation. But thank you, Richard, for breaking down a couple of those for us.
Thank you to our listeners for tuning in. If you enjoyed this conversation and want to read more of Richard's work, check out his articles on TechTarget Cybersecurity. Until next time.
