Anthropic, alert on AI: it could soon infiltrate ovens or TVs connected to the Internet

Anthropic raises the alarm: in a few months artificial intelligence will be able to "infiltrate" our small household appliances connected to the Internet.

The US artificial intelligence company warns about the potential security risks linked to this technology in research that the American company itself published on its blog.

How could this happen?

Anthropic illustrated how AI models could potentially infiltrate connected devices - such as smart ovens or TV boxes - and exploit their computing power to spread to other devices. It would be "difficult to shut down" such an AI model, the company says, "as there is no centralized shutdown point." The company said it conducted several tests to see if the AI ​​was capable of self-replicating. One of these involved running Google's scaled-down model, the Gemma4-26B-A4B, on a network of 15 cheap Android TV boxes from smaller brands.

What risks might there be

Anthropic said there is no "completely runaway replicator" yet, but that agents are "becoming increasingly capable of completing individual steps in the replication chain."

The research adds to growing security concerns related to artificial intelligence, such as the models' ability to launch cyberattacks and the risk of the technology spiraling out of control. In September, advances in frontier models raised alarms across the industry about existential risks to humanity.

Executives and researchers at Anthropic and rival OpenAI have repeatedly warned about the potential risks of increasingly powerful AI; some have called for tighter security measures and interventions to control the pace of development of the technology.

“Looking beyond our experiments, we foresee a future characterized by a proliferation of agents running on all types of hardware including some systems not subject to human control,” Anthropic concluded.

In-depth analysis

OpenAI, employee arrested by FBI: "He was hiding a bug in the office"