The key facts

  • Anthropic demonstrated in a controlled experiment that an AI model can spread between cheap Android devices: in 4 out of 5 attempts it infected all 8 target devices.
  • According to Anthropic's estimate, an 'AI worm' could establish itself on a new device within one to two days, while smaller AI models may reach the required capability level 'later this year'.
  • In France, supermarket refrigeration systems were attacked (temperatures up to 50 °C); CERT-FR confirmed the compromise of Danfoss systems, no link to AI worms has been established.

Anthropic's warning about AI worms

Anthropic has published a report warning that malware based on artificial intelligence could soon penetrate connected household devices, copy itself onto other devices and activate there. According to a study by the industry association Bitkom, about half (52%) of households in Germany already use 'smart' applications — robot vacuum cleaners, voice assistants, heating control and internet-connected TVs — and this share is growing.

How AI worms differ from conventional ones

Classic computer worms that copy themselves from machine to machine have existed for decades and follow a predefined sequence of instructions, exploiting known vulnerabilities. AI worms would be considerably more sophisticated: they could autonomously solve tasks, adapt to different devices (for example, a different operating system or limited memory), independently learn to operate on new devices and circumvent new security updates.

Experiment on cheap hardware

Anthropic researchers tested whether AI models can run on inexpensive hardware: they distributed an AI model across 15 cheap Android TV boxes, each of which performed part of the computations. This showed that some AI applications can run without a large data center. In a controlled test environment with eight additional simulated Android devices, the model managed to penetrate one of them, copy itself and infect the rest: in four out of five attempts it reached all eight targets, and in the remaining one, seven of eight. According to Anthropic developers' estimate, such an 'AI worm' could establish itself on a new, previously unknown device in roughly one to two days.

Risk of a botnet made of household devices

Particular concern is raised by the possibility of combining infected devices into a large botnet: the more devices are compromised, the faster the worm could spread, and the malware would no longer depend on a central server that can simply be switched off. The worm could continuously learn as it spreads and possibly move to entirely different types of devices with different software. The decentralized structure would make containment difficult, despite the possibility of disconnecting devices from the network and installing patches. Anthropic compares such a development to a slowly spreading digital slime mold. The researchers deliberately simplified access to the test devices and did not use serious digital protections; however, the company believes that smaller AI models may reach the necessary capability level 'later this year' to spread in the real world.

Attack on supermarkets in France

The possible consequences of attacks on connected control systems are already visible in the example of an incident in France: overnight into Monday, attackers interfered with the operation of refrigeration and freezing systems at numerous supermarkets. The temperature in the affected units rose to 50 degrees Celsius, food spoiled, and in some cases there were fires and technical damage. France's cybersecurity agency CERT-FR initially confirmed the compromise of remote control systems made by the Danish manufacturer Danfoss, and later extended the warning to systems from other manufacturers. The investigation is ongoing. It was reported that some of the protective software was outdated — this may have been the entry point for the attackers.

No link between the attacks and AI established

Who was behind the attacks in France and what the motives were remains unclear. There is also no evidence yet that a self-replicating AI worm was used in the attacks on the refrigeration systems. Nevertheless, the Anthropic study shows how much AI could facilitate the creation of malware, potentially making it possible to attack connected devices with serious consequences and without significant computing resources. Fearing repeat attacks, CERT-FR recommended that companies regularly update connected devices, install the necessary security patches and, where possible, completely disconnect unneeded internet access.