Anthropic: Small AI Models Could Infect Smart TVs and Robot Vacuums Within 6–12 Months

Small AI models could start spreading on their own across poorly secured consumer devices within months: smart TVs, robot vacuums and other "smart" electronics. That is the conclusion Anthropic researchers reached. In their tests, one model already completed the entire infection chain. CEO Dario Amodei warns that within 6–12 months such a swarm could build a persistent botnet the size of the whole internet, causing hundreds of billions of dollars in damage.

What the tests actually showed

The public summary of the research offers few details. It is unclear which model passed the test, on which devices, and through which vulnerabilities. Those details are best sought in the report itself. But the main claim is stated clearly: this is a full infection chain, not hints given to a human. The model goes from target to persistence in the system to further spread, with no operator holding its hand.

For the security industry, this is a shift from theory to practice. Until now, talk of "self-replicating AI" remained a future scenario. Now Anthropic has a laboratory result, and the company itself speaks of a horizon of months, not years.

Why small models in particular are dangerous

Large frontier models live in data centers, and they cannot be copied onto a TV. Compact models are a different matter: they need far less memory and compute, and in principle can run on modest hardware. This is our inference from the logic of the problem, not a quote from the research, but it explains the focus on small models.

The targets are not chosen at random either. Cheap consumer gadgets get poor updates for years, or none at all. Default passwords, exposed services and ancient firmware are commonplace. There are millions of such devices, and their owners rarely look at the settings.

A botnet that can think

The closest historical analogue is Mirai. In 2016, this botnet of cameras and routers helped knock out some major services by attacking the DNS provider Dyn. Mirai followed a rigid script: it cycled through known passwords and acted on predefined rules. If a device did not fit the template, it was left untouched.

A model with its own reasoning can adapt to a specific firmware and look for workarounds where a script hits a wall. That is exactly what Anthropic fears. Amodei's warning about a "persistent botnet" and hundreds of billions of dollars in losses should be read as a scenario, not a measurement. A company that trades on its reputation as a responsible developer has an interest in such forecasts. But the fact that the full chain has already been reproduced in a test is not a forecast.

Key figures from the statement

• Threat horizon: 6–12 months, according to Amodei.
• Potential damage: hundreds of billions of dollars.
• Type of targets: smart TVs, robot vacuums and other poorly protected consumer devices.
• Test result: at least one model completed the entire infection chain.
• Class of models: small, able to run on limited hardware.

What owners and manufacturers should do

Until the full details of the research are published, basic hygiene remains the best defense. Owners should change default passwords, install the latest firmware and disable any remote access they do not need. It is better to move these devices onto a separate guest network, so that a TV or vacuum cannot see laptops and network storage. If the manufacturer has stopped releasing updates, it makes sense to replace the device.

Manufacturers will have to rethink the "ship it and forget it" approach. They need unique passwords on every device, regular automatic updates and a minimum of open services. Corporate security teams should inventory everything connected to the office network, from smart displays in meeting rooms to cameras. For an attacking agent, these are the softest spots in the perimeter.