Self-Replicating AI Warning: Is Your Smart Home at Risk?

The cheap TV box under your television has made its way into an AI safety paper. On October 12, 2026, Anthropic warned that small models may be months away from self-replicating AI that can copy itself onto inexpensive connected devices. No current model can do this reliably on a smart TV, and no home infections of this kind have been reported.

So, should you unplug everything? The research gives you reasons to pay attention, with clear limits on what has actually been demonstrated. The router and device protections used against existing botnets remain your best defense. Start with the key facts.

Item	Detail
Who	William Zhou and Daniel Freeman, Anthropic Frontier Red Team
Published	October 12, 2026
Claim	Small models may be months from independent replication on cheap devices
Devices named	Android TV boxes, smart TVs, fridges, ovens, vacuums, lawnmowers and some cameras
Expected timing	Around the end of 2026, based on a rough estimate
Active attacks	No home AI worm infections reported
Best defense	Secure your router and connected devices

What Anthropic Found About Self-Replicating AI

Anthropic tested whether AI could run and reproduce across hardware far smaller than a datacenter. For you, the relevant detail is that a model does not need to fit inside one appliance. Researchers showed it could run when divided among several inexpensive devices. Here are the numbers that explain both the progress and the limits.

• Hardware coverage. A model with nine billion parameters managed to run AI computations on eight of the nine hardware categories tested. That demonstrates compatibility, rather than reliable replication.
• Incomplete replication. Models with 30 billion parameters or fewer could handle many separate tasks, but made errors when connecting those tasks into a complete process.
• Simplified success. GLM-5.2 reached all eight cloud nodes in four of five trials. In the remaining trial, it reached seven. Gaining access was deliberately easy, so you should read this as an upper limit on capability.
• Cheap hardware demonstration. Fifteen off-brand Android TV boxes, costing about $55 each, ran a divided 26-billion-parameter model at roughly 0.39 tokens per second. It worked, slowly.
• Estimated timing. Researchers projected that a small model could reach the required capability around the end of 2026, and small replicators could come after that. They described that timetable as rough.

You see, running AI on a box and having AI spread reliably are separate achievements. The study provides evidence of progress toward the second, without proving that your television can host a reliable replicator today.

How an AI Worm Differs From an Ordinary Botnet

An AI worm, in this research, is a self-spreading program driven by an AI model. Its proposed advantage is the ability to work out unfamiliar devices as it goes. Copies distributed across devices would also lack a central server you could simply switch off. To understand the difference, compare it with the botnets already discussed in the research.

Trait	Classic botnet	AI worm as described by Anthropic
Device targets	Kimwolf was built for particular device types	Could operate across unfamiliar hardware
Getting access	Mirai relied on unchanged manufacturer passwords	Would scan, gain access and deliver its payload
Adapting	Uses instructions written for its targets	Could reason about a new device and adjust
Running on devices	Kimwolf reportedly infected over two million devices, mostly off-brand Android TVs	Could divide an AI model among several devices with limited memory

The term has another meaning, too. In 2024, researchers built Morris II to spread through AI email assistants using malicious prompts hidden in retrieved material. Their tests demonstrated click-free propagation, spam and contact-data theft. That worm travelled through AI applications. Anthropic’s research examines a model copying itself onto connected hardware, so you should keep the two types separate when reading headlines.

Which Home Devices Are on the List

Your device does not need to be a powerful computer to enter this discussion. Cheap boxes commonly have 1 to 4 GB of memory, enough for part of a model when several devices share the work. Researchers identified online devices with a history of weak security as possible early targets. Here is what each category contributes to the concern.

• Off-brand Android TV boxes. These inexpensive devices supplied the hardware for Anthropic’s demonstration and were the main devices reportedly infected by Kimwolf.
• Smart TVs. Their computing resources make them relevant, though no current model is both small and capable enough to replicate reliably on one.
• Smart fridges and ovens. Connected appliances were named as devices that could hold a portion of a smaller model.
• Robot vacuums and lawnmowers. Researchers included these connected machines among possible hosts for model fragments.
• Some cameras and conferencing units. Versions with AI chips may have the resources to run part of a model.

Anthropic explicitly says the chips it named are not themselves security flaws. Your appliance appearing on this list does not establish that it has been compromised.

How Worried Should You Be Right Now

You have a research warning to assess, rather than a reported outbreak in homes. The experiments show useful progress, but their conditions matter just as much as their results. Calling the warning fear mongering dismisses those results, while treating it as an active attack ignores the caveats. Keep both sides in view.

Reasons for calm

• The capability gap remains. Researchers found no current model that combines the necessary ability with a size suitable for reliable replication on a smart TV.
• Access was made easy. The successful cloud test used an unauthenticated port, making the intrusion task trivial. That limits what you can conclude about protected networks.
• Setup took many hours. Estimates for real TV boxes ranged from 28 to 150 hours, depending on the model. The demonstration was far from quick.
• Spread would be slow. Anthropic expects a replicator’s population to double over days. Small models also performed poorly at targeted phishing.
• Today’s threat has limits. AI safety researcher Roman V Yampolskiy said this is not an existential threat today, although he finds the trend concerning.

Reasons to pay attention

• The estimate is near. The end-of-2026 projection gives you a reason to follow small-model progress, while remembering that it is an estimate.
• Experience improved results. After notes passed through six generations, one model’s setup success climbed from 45% to a range of 61% to 84%.
• Shutdown could be difficult. A replicator spread across devices would have no single central server to disable.
• Other lab work supports the concern. Palisade Research showed models copying themselves between deliberately vulnerable computers in May 2026. Those controlled tests add evidence, with the same need to account for easier test conditions.

Steps to Lock Down Your Smart Home Today

You do not need to wait for an AI worm to improve your home network security. CISA and the FTC already recommend measures against unauthorized access and existing malware. Anthropic’s paper does not provide a consumer checklist, so the steps below come from those agencies’ guidance. Start with your router, then check the devices connected to it.

1. Replace the router’s default admin login. Use a long, random and unique username and password. Keep the admin password different from your Wi-Fi password.
2. Strengthen your Wi-Fi password. CISA recommends five to seven unrelated words totalling at least 16 characters. Choose a passphrase you have not used elsewhere, and keep personal information out of your network name.
3. Check your Wi-Fi encryption. Use WPA3 Personal where available, or WPA2 with AES. If your router offers only WEP, WPA or WPA2 TKIP, CISA recommends replacing it.
4. Update the router’s firmware. Install available updates and enable automatic installation if offered. Register with the manufacturer for update notices, or ask your internet provider whether it manages these updates.
5. Disable unnecessary access features. Turn off remote management, WPS and UPnP. CISA notes that UPnP can help malware spread. If setup requires it, enable it briefly rather than leaving it on.
6. Give smart devices a guest network. Move devices that only require internet access onto that network. CISA recommends this separation so they cannot reach your other devices or the router’s controls.
7. Keep the router firewall enabled. The FTC recommends checking that it is switched on. Treat it as one part of your protection, alongside updates and the other steps here.
8. Log out after making changes. Leave the router’s admin account signed out when you finish. CISA also recommends keeping the router physically secure.
9. Check every connected device. Apply available updates and secure each gadget. The FTC warns that an unsecured device can provide attackers with a way into your network.

Want more help? Our guides to fixing a smart plug that won’t connect to WiFi, spotting the signs your phone is hacked and building a smart home for your kids walk through related checks.

Buying Smarter Connected Gear

Update support and secure defaults deserve attention whenever you buy connected equipment. Use these checks before adding another device to your home.

• Check for the U.S. Cyber Trust Mark. Announced in January 2025, this FCC-run programme covers consumer connected products, including cameras, appliances and baby monitors. Participating products undergo testing against NIST criteria through accredited laboratories.
• Read the QR code information. The label links to a registry with the product’s support period, automatic-update details, password-change instructions and secure setup guidance. You can assess those details before relying on the device.
• Be cautious with no-name Android TV boxes. Kimwolf reportedly infected more than two million devices, mainly off-brand Android TVs. Anthropic also demonstrated model operation on inexpensive off-brand boxes. Neither finding establishes that every cheap box is compromised.
• Ask about updates and defaults. Check how long the maker provides support. The UK ban on default connected-device passwords is one example cited in the research of requiring stronger security at the start.

What ISPs, Makers and Regulators Are Being Asked to Do

Your router checks matter, but Anthropic also proposes action beyond individual homes. Its defensive goal is to reduce replication until each infected device generates fewer than one additional infection on average. The proposals assign different jobs to providers, manufacturers and policymakers, as this table explains.

Who	Suggested action
Internet service providers	Watch for unusually large model transfers and abnormal traffic between devices
Network defenders	Use decoy systems, known as honeypots, to help detect replication activity
Internet service providers	Isolate infected devices and notify customers so they can address the infection
Device makers	Supply products with secure defaults rather than leaving protection to the buyer
Regulators and policymakers	Support requirements for connected-device security, with the UK password rules as an example

Customer notification has evidence behind it. In the Mirai research cited by Anthropic, isolation plus notification cleared 92% of infections within 14 days, compared with 74% without notification. You should read those figures as evidence from Mirai, rather than results from an AI worm outbreak.

What Happens Next

Watch for evidence that small models can complete replication reliably around the end of 2026. That date is a rough research estimate, and no home AI worm infections have been reported. For now, you can act on the security guidance already available. Start with your router login, updates and connected devices.

You see, checking passwords and update support gives you practical work to do while the research develops. We will update this guide if actual home infections are reported. Until then, keep lab demonstrations, predictions and confirmed attacks separate when you read the next headline.

FAQs