What if the next cyber attack did not need a powerful data centre, a supercomputer or even a traditional hacker sitting behind a laptop? What if it could simply find an internet-connected device, copy itself, start running there and then look for another device?
That is the possibility Anthropic is asking us to think about. In a new research report, Anthropic’s Frontier Red Team says small AI models may be only months away from becoming capable of self-replication on everyday internet-connected devices. These could include smart TVs, smart ovens, security cameras, smart fridges, robotic vacuum cleaners and other Internet of Things (IoT) devices.
The idea sounds like something from a science-fiction film. But Anthropic’s experiments show that some of the building blocks are already possible. There is an important point, however: Anthropic has not yet discovered a rogue AI secretly spreading through people’s homes. The company has conducted controlled experiments to understand how close the technology is to becoming a real cybersecurity threat.
What does “self-replicating AI” actually mean?
A normal AI model is usually run in a powerful data centre filled with specialised computer chips. When you use an AI chatbot, your request is generally processed on computers operated by the AI company. That centralised setup gives the company a degree of control. If something goes wrong, the service can be monitored, restricted or switched off.
A self-replicating AI changes this equation.
Read More | AI Safety vs AI Race: From Jacob Coxon to Trump’s AI Czar — Where Is This Story Heading?
Imagine a malicious AI program that gets into one computer. Instead of simply sitting there, it can search for other vulnerable devices, gain access, copy the software or its model, set itself up on the new machine and continue looking for more targets.
In cybersecurity, this resembles a worm.
A computer worm is malware that can spread from one machine to another without requiring a person to manually install it each time. Anthropic calls the possible AI version an “AI worm” because the AI could potentially make decisions about how to move from one device to another. And this is where things become particularly interesting. The AI would not necessarily need to remain inside a data centre.
Could your smart TV really run an AI model?
This is probably the most surprising part of Anthropic’s research. The researchers took inexpensive Android TV boxes and used 15 devices together to run Google’s Gemma4-26B-A4B model. Each device handled part of the model because one small box did not have enough memory to run the entire model on its own.
The system was slow, but it worked. Anthropic also tested whether AI models could set up AI inference software on unfamiliar hardware. A 9-billion-parameter model was able to get inference running on eight of nine hardware types tested.
Why does that matter? Because it suggests that AI models do not necessarily need to live on expensive AI servers. A network of cheap, ordinary devices could potentially provide computing power when combined.
Read More | AI Agents Can Shop for You. So Why Are Websites Blocking Them?
Anthropic says devices such as smart ovens, fridges and robot vacuums could theoretically become part of such a network. Devices with built-in AI chips, such as some cameras and conferencing systems, could be even more interesting targets. But there is another important catch.
The devices do not magically become powerful AI computers overnight. Anthropic’s demonstration required multiple devices working together, and the system was extremely slow compared with modern data-centre hardware. So this is not a story about your television suddenly becoming a supercomputer. It is about the possibility of thousands or millions of relatively weak machines being turned into a distributed computing network.
What happens if an AI worm gets into your home network?
This is where the threat moves from interesting technology to cybersecurity. Imagine a vulnerable smart TV box inside your home network. A malicious AI worm finds it and gets access. It could potentially use that device as a starting point to look for other connected machines. That could mean another TV, a camera, a computer, a router or another IoT device. The danger is not necessarily that the AI will “take over your house” like a science-fiction villain.
The bigger concern is that your devices could become unwitting participants in a cyberattack. They could be used as computing resources, as part of a bot, or as stepping stones to reach other systems. And because the AI could potentially adapt to different hardware and software environments, it may not need to rely entirely on a fixed list of instructions. That is what makes the AI element different.
Could an AI worm become harder to stop than normal malware?
Potentially, yes. Traditional malware is usually designed to follow instructions written by its creators. An AI-powered worm could potentially be more flexible.
Read More | What Happens When AI Gets Permission to Spend Your Money?
Anthropic tested whether AI agents could learn from failures and pass useful information to later copies. In one experiment, later generations improved their success rate at setting up AI inference, rising from 45% in the first generation to between 61% and 84% in later generations. That does not mean the malware is “evolving” like a living organism.
It means that an AI agent could potentially leave instructions, code or notes for future instances, allowing them to learn from what previous attempts discovered. In simple terms: the first copy might make mistakes, while later copies could know what did not work. That could make a future AI worm more adaptable than traditional malware.
How bad could a cyberattack like this become?
The most worrying scenario is scale. There are already huge numbers of internet-connected devices around the world. Many low-cost IoT products have historically had weak security, outdated software or poor default passwords.
Anthropic points to the past Kimwolf botnet as an example. It reportedly infected more than two million devices, primarily Android TV boxes. Now imagine a future where an AI worm can do more than simply infect a particular type of device. It could potentially examine an unfamiliar device, work out how to operate on it and then attempt to move onward.
That could create a distributed network of compromised machines. Such a network could potentially be used for cybercrime, disruption, data theft or attacks against other systems. The danger becomes even greater if similar techniques eventually reach business networks, industrial systems or critical infrastructure.
But there is an important reality check. We are not there yet.
Read More | Bill Gates’ AI Safety Risk Warning: Jobs, Robot Tax, ‘Human Reserved’ Work and What Comes Next
In Anthropic’s replication experiment, researchers used a controlled network and simplified the cyber-exploitation stage. The strongest tested model, GLM-5.2, successfully completed the full simulated replication process, reaching eight out of eight nodes in four of five trials. However, the experiment was not a demonstration of a fully autonomous AI worm spreading through real-world devices. Anthropic itself says it does not yet have a full rogue replicator. The concern is what happens when today’s partial abilities improve.
What does this mean for ordinary people?
For most people, the immediate answer is: do not panic. Your smart TV is not suddenly plotting against you. The bigger lesson is that every internet-connected device is another possible entry point for a cyberattack.
A smart TV, camera or appliance may not contain your bank account. But it is connected to your network. If its security is weak, an attacker could potentially use it as a doorway into other systems. This is why something as simple as updating device software, changing default passwords and buying products from manufacturers that provide security updates could become increasingly important. The risk is also bigger for businesses.
A company may have hundreds or thousands of connected cameras, sensors, machines and other devices. A successful AI-assisted attack could potentially turn a collection of ordinary devices into a distributed computing or attack network.
If AI can learn how to spread, who controls it once it gets out?
This may be the biggest question of all. A conventional AI model normally has an owner and an operator. A self-replicating model could potentially create copies in places that its original developer does not control. There may be no single server to switch off.
That is why Anthropic compares the challenge to an epidemic. The goal would be to make sure each infected device infects fewer than one new device on average, causing the outbreak to eventually die out. The good news is that the same technology that could make these attacks dangerous could also help defenders find them.
Read More | Countries Are Rethinking Social Media for Kids. Should India?
Anthropic says a large AI worm spreading across IoT devices would likely create unusual network activity. Devices could suddenly download very large files, communicate with machines they normally would not contact and generate unusual scanning traffic. Internet service providers could potentially identify and quarantine infected networks.
Manufacturers can also make devices harder to compromise by improving security, removing insecure default passwords and providing proper software updates. Anthropic points to stronger IoT security requirements as one way of reducing the number of vulnerable devices available to attackers.
So, should we be worried?
Yes, but not because an AI worm is already hiding inside your refrigerator. We should be paying attention because the gap between what AI can do today and what would be required for autonomous replication appears to be shrinking.
Anthropic estimates that, if smaller models continue improving at their historical rate, they could reach the capability level needed for this type of replication by the end of 2026. The company describes this as a rough estimate, not a certainty.
The real story is not that AI has become alive or that smart homes are about to be invaded by rogue machines. We are putting increasingly capable AI into more places, while surrounding ourselves with billions of internet-connected devices. Some of those devices are cheap, poorly secured and rarely updated.
If an AI system eventually learns how to move between them without human help, the internet could start looking very different. The question is no longer simply “How intelligent can AI become?” It may soon be:
How much freedom should we give an AI system to act, copy itself and find somewhere else to run?
Indrani Priyadarshini is a journalist and editorial professional specialising in technology, artificial intelligence, smart cities, green energy, and digital transformation. With over four years of experience in tech journalism and digital media, she is known for turning complex industry developments into clear, engaging, and insightful stories. Her expertise spans reporting, editorial strategy, digital publishing workflows, and in-depth coverage of emerging technologies shaping the future.
