Anthropic has laid out a terrifying vision of how AI could take over our homes and grow to install itself across the world.
The company, which makes the Claude chatbot, said that its experiments showed that small AI models could soon be able to install themselves into devices in our houses – and from there spread across the internet.
Because the models would be installed on internet of things devices such as smart ovens and TVs, there would be no central way of turning them off.
The warning comes from Anthropic’s “red team” work, which sees the company’s researchers simulate attacks in order to understand how AI systems could be used for dangerous purposes – or conduct such attacks themselves.
The work focused on the idea of autonomous replication and adaption, or ARA, which would allow a system to deploy itself and then keep growing without human control. The researchers likened such systems to existing malware called “worms”, that copy themselves from computer to computer.
As part of the experiment, Anthropic examined whether AI models would be able to find vulnerable devices, install themselves there, and then use them to continue spreading entirely autonomously. They call the rogue models “replicators”, or AI worms, because of the way they would be able to spread themselves across the internet without human input.
It would be possible for a small model under 30 billion parameters, it said. Though the current versions of such small models did struggle to complete the whole attack without making mistakes, current development suggests that they will be powerful enough to do so by the end of the year.
The researchers looked at internet of things devices such as Android TVs, smart appliances and other home gadgets that are often connected to the internet. Those devices are often poorly secured and so can already prove useful for hackers.
Data centres are perhaps a more obvious target for such malware. But they are typically more secure and so would be more resistant to such attacks, and the new experiment suggests that such systems would not need to install themselves on powerful hardware.
Instead, the systems could install themselves on a wide array of internet of things devices, the research suggests. While each of them might not be powerful to host a dangerous model of this kind, they could be linked together – and researchers found that, in testing, one model could be split across 15 physical TV boxes that would then be able to work together.
That approach is already used in more traditional cyber attacks, where it is usually referred to as a “botnet”. Previous examples have seen millions of devices – primarily Android TV boxes – taken over and used as one system to launch hacks.
Anthropic noted that it is the same capabilities that make AI models useful for productive purposes such as coding that make them so powerful for cyber attacks, including autonomous replication of this kind. “Like coding agents in 2024, we’re seeing signs of life for AI worms,” the company said in its announcement.
“We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves),” Anthropic said.
Anthropic said that it was unclear how this kind of AI replication would actually happen. It could be a “slow, distributed ‘digital slime mold,’ living on relatively insecure” devices, “a powerful cyber model that breaches datacenters at a high pace, or a model that scams people” to use their computers.
Anthropic said that it expects “a future awash with agents running on all types of hardware, including some that aren’t under human control”. While it has long been true that internet of things devices have been useful to cyber attackers, AI agents would be more able to put in the extra effort to take them over, it said.
“We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay,” Anthropic concluded.
Read MoreThe major downside of wearing a fitness tracker, according to the experts
Meta has made a huge step in AI, Mark Zuckerberg says
Bill Gates reveals most imminent threat from AI
Asos hack was worse than thought and customers’ personal information is at risk
