Anthropic warns of AI worms on IoT devices: advice for small offices

What happened

Anthropic Frontier Red Team, the Anthropic team that studies the potential risks of AI models, published a study on 12 October 2026 on whether small AI models could copy themselves onto cheap, internet-connected devices. Its conclusion: such models may be only months away from being able to copy themselves onto commercial IoT devices, such as smart ovens and cheap unbranded Android TV boxes, without needing a data centre. The study was also widely covered in the media.

The researchers treat this activity as autonomous replication and adaptation (ARA) in four stages: finding targets, gaining access, launching the model on the new device and spreading further. The study focused on the stage of launching the model. Key findings:

Google's Gemma4-26B-A4B model was split across 15 cheap unbranded Android TV boxes (about USD 55 each) and ran at 0.39 tokens per second.
A 9-billion-parameter model was able to launch itself on 8 of the 9 device types tested.
Of the 11 open models tested, only GLM-5.2 achieved full propagation in a simulated environment, reaching all 8 nodes in 4 of 5 attempts.

An important caveat: the break-in step in the test was deliberately made easier, because the ADB port had no authentication, which is why the researchers call the result an upper bound. This is a capability test in a controlled environment. A fully autonomous AI worm on real devices was not demonstrated, and none has been found in the wild. The researchers estimate it could become possible by the end of 2026 if models keep improving at the current pace.

What this means for companies in Latvia

According to researchers William Zhou and Daniel Freeman, the first target could be home and small-business IoT devices, because they are connected to the internet and have historically had minimal security. This applies to small offices in Latvia as well. In a typical office, far more than just work computers is connected to the internet: a router, security cameras, a TV box in the meeting room, a printer, and sometimes a coffee machine or other appliances. Many of these devices were set up once and have not been touched since.

The researchers note that such a spreading tool would be hard to stop, because there is no single central point to switch off. At the same time, it could be slow, as its doubling time is measured in days. Unlike ordinary malware, it could work out unfamiliar devices on its own rather than relying only on pre-prepared instructions.

However, the vulnerability itself is not new. Earlier IoT botnets, such as Mirai, spread by exploiting manufacturers' default usernames and passwords. In the study's test, too, an unprotected ADB port made the path easier. In other words, the door through which future threats could come in is the same one many offices already leave open. There is no need to panic, but there is good reason to sort out the basics.

How to apply this in practice

The steps below are not findings of the study, but generally recognised good practice that AI-Lab Space recommends to every small business, regardless of the AI topic.

Know what is connected to your network. Log in to your router's admin panel and look at the list of connected devices. Often there are devices nobody remembers any more.
Change factory default passwords. This applies to the router, cameras, printers and all smart devices. Each device should have its own strong password.
Update software regularly. Turn on automatic updates where possible, and check the other devices every couple of months.
Separate cheap smart devices. Keep TV boxes, cameras and appliances on a separate network, for example a guest network, apart from work computers and company data.
Drop what you do not need. If a device does not need to be online, disconnect it. If a device is no longer in use, remove it from the network.

What does this look like in real life? Before: the meeting-room TV box, a camera and the accounting computer are on the same Wi-Fi network, the camera still has its factory password, and the TV box has not been updated since purchase. After: smart devices run on a separate network, passwords have been changed, updates are switched on, and a device that is no longer used has been disconnected. Work data is no longer on the same network as the cheapest and least protected devices.

Security when adopting AI

This study is also a reminder for those introducing AI tools or AI agents in their business. Security must be built in from the very start, not added later:

Clear access rights. An AI agent has access only to the systems and data needed for the specific task.
Data under control. The company knows where data is stored, who can access it and how it is processed.
A human approves actions. Important actions, such as payments, sending emails to customers or changes to systems, are approved by a person.

Where to start

You do not have to rebuild the whole office in a single day. One calm afternoon is enough:

Write down a list of all internet-connected office devices.
Note which devices still have default passwords, and change them.
Create a separate network for smart devices and move them there.
Disconnect whatever is not needed.
Set a calendar reminder to check for updates regularly.

The Anthropic Frontier Red Team study shows the direction in which technology is developing, not an event that has already happened. There is no need to panic, but basic hygiene should be followed now, because it protects against today's threats and prepares you for future ones too.

If you want to introduce AI in your business safely, with clear access rights, controlled data and human approval for the most important actions, contact AI-Lab Space. We will help you assess your current situation and build a solution where security is planned from the start.