Preparing for an AI ‘kill switch’

Fears that AI could escape human control have heightened calls for a ‘kill switch’, which would require companies developing frontier models to cut off access to their tools.

Politicians in the US have been putting forward a series of legislative proposals in an attempt to rein in AI models that could go rogue.

The AI Kill Switch Act was introduced in late July with a view to forcing developers to throttle, suspend, or shut down AI systems at risk of causing catastrophic damage.

“As a computer science major, I am very aware of the dramatic possibilities – both good and bad – that AI presents,” Democratic congressman Ted Lieu said in a press release. The bill had the support of both parties.

The Stop Rogue AI Act, another bipartisan bill, was introduced on September 15. It would guarantee the US National Institute of Standards and Technology develops standards, guidelines, and best practices for controlling AI agents

“Right now, AI agents are running loose in our networks, and nobody can see them or verify who built them, making it increasingly hard to stop them,” Democratic congressman Josh Gottheimer said in a press release.

An AI Emergency Button Act was introduced by Republican congressman Tom Kean Jr on September 24. “While AI can be a helpful tool, it is essential that humans remain in control of complex artificial intelligence systems,” he said in a press release.

Legislation is already facing hurdles, though. An attempt to fast-track the passage of a separate AI Emergency Button Act was blocked by the Senate on September 16. It could be months before President Trump actually signs a bill into law.

The operational risks of a ‘kill switch’

While legislation is well-intentioned, there is an argument to be made that any ‘kill switch’ would be difficult to implement from a frontier model perspective.

“If something goes wrong, for example, an AI attack on national infrastructure, you’d want to be able to roll it back and switch it off. In reality, though, AI exists at every level. It’s on our mobile devices, across physical networks, in data centre networks, on our laptops. You'd have to find all of it before you could shut it down,” Charlotte Wilson, head of enterprise at Check Point, tells ITPro.

Gary Barlet, public sector CTO at Illumio, adds that a ‘kill switch’ could cause systemic risk given how embedded AI has become in business operations.

“If a model suddenly becomes unavailable, those workflows could be disrupted overnight. Even intended as a safety measure, it could become a source of uncertainty, disruption, and economic harm if it's triggered too aggressively or without clear standards,” Barlet says.

Darren Guccione, CEO and co-founder of Keeper Security, points out that any ‘kill switch’ would not make AI secure by itself and definitely wouldn’t address every risk. However, it would help to establish an emergency baseline that frontier model developers have to meet.

Preparing for a shutdown

The question for organizations is whether they’re prepared for developers having to design a ‘kill switch’ into their frontier models – and how they would react to it.

Guccione says that large companies generally have a bigger pool of resources to invest in a wide range of models. However, it’s the smaller companies that are likely to be dependent on the same frontier models, putting their operations at greater risk of being impacted by any switch-off.

Smaller companies are also less likely to invest in their own AI safety mechanisms. Kiteworks’s annual data security and compliance risk report, published in July, found that the ‘kill switch’ gap is widening. Researchers at the software development company had forecast that 60% of organizations in 2026 wouldn’t have a tested capability that can shut down an active AI agent in real time, but the actual figure was 79% among 459 security and compliance professionals surveyed.

In the event a shutdown suspends a model that workflows are built around, the onus is on organizations to keep whatever depends on the model functioning, Guccione stresses. “Continuity planning needs to treat that as an operational imperative in its own right."

It may seem obvious, but to prepare for a shutdown, organizations need to improve their visibility by establishing an inventory of AI models and their dependencies.

“They should identify where AI models are embedded in business-critical processes, which providers those processes rely on, and what data and systems are connected to them,” Guccione continues.

From there, organizations need to put a fallback plan in place for their most important business processes. Where possible, they should “avoid unnecessary dependence on any single provider” and seek out alternative models to prevent single-model lock-in, Barlet advises.

A ‘kill switch’ would increase confidence in AI

While unexpected shutdowns caused by a ‘kill switch’ could scare organizations into stopping using models, its introduction would likely drive AI adoption in the long run.

As Barlet explains: “There's a tendency to assume regulation slows adoption, but uncertainty is often the bigger obstacle. When organizations understand the security requirements, testing standards, and governance obligations, they're more willing to invest. The right safeguards can increase confidence.”

Ultimately, a ‘kill switch’ would mean that organizations don’t have to rely on vendor assurances about AI safety, Guccione adds. “It would shift trust from claims made by individual vendors towards controls that organizations can actually verify.”