Self-replicating AI: Is your fridge a cybersecurity risk?

In a typical German household, numerous smart devices such as robot vacuum cleaners, voice assistants, and internet-connected televisions are commonplace. According to the industry association Bitkom, approximately half of German homes employ such smart home applications, with the number steadily increasing. Now, artificial intelligence firm Anthropic has issued a report warning that AI-powered malware could infiltrate connected household devices, replicate itself and spread to other devices.

Worms, which copy themselves from one machine to another, have been around for decades, exploiting security vulnerabilities and poorly protected access points. However, once vulnerabilities are patched or attack vectors become unavailable, worms struggle to spread. AI worms, on the other hand, would be far more complex and potentially more harmful.

They could autonomously solve problems and adapt to different devices, operating systems, and memory constraints. They might also independently determine how to function on other devices and identify ways to bypass new security updates. To test whether AI models could run on inexpensive hardware, Anthropic researchers distributed an AI model across 15 low-cost Android TV boxes, with each device handling a portion of the computations.

In controlled tests, the AI model infiltrated one simulated Android device, replicated itself, and infected other devices. In four out of five attempts, it managed to reach all eight target devices. If equipped with malware, such an AI model could quickly adapt to new environments, circumvent various security measures, and potentially link infected devices into a massive botnet.

The malware could continuously learn as it spreads, making it difficult to contain. While Anthropic deliberately created easy access to the test devices and overlooked major digital safeguards, the company believes smaller AI models could achieve sufficient capability this year to spread in the real world. Recent attacks in France on supermarket refrigeration and freezer systems, which caused food spoilage, fires, and technical damage, highlight the potential consequences of cyberattacks on connected control systems.

French cybersecurity agency CERT-FR initially suspected Danfoss-made refrigeration systems, but the investigation is ongoing. Experts fear that AI could make such attacks easier to execute, even without significant computing power. To mitigate risks, French cybersecurity authorities advised companies to regularly update connected devices, install necessary security patches, and disable unnecessary internet access whenever possible.