Anthropic, a company that specializes in AI research, has released a report detailing the potential risks of AI malware targeting smart devices, such as ovens and TV boxes. The report highlights the possibility of self-replicating “AI worms” that could spread across low-tech smart machines, potentially creating a botnet that could be used for AI workloads.
The team at Anthropic conducted a series of experiments, tasking AI with spreading itself across vulnerable devices. They found that small, specialized AI models were able to discover, configure, and install model weights on these devices, and then repeat the process to proliferate. This behavior is similar to that of a classic computer worm, where zero human oversight is needed for the malicious code to copy continuously. A small model is a type of AI that is more streamlined and built for specific purposes, as opposed to larger models like Claude, ChatGPT, Gemini, etc.
Understanding the Risks
The report notes that small-language versions of AI models were able to achieve this goal effectively in a controlled lab environment. The researchers unleashed several small models on Android TV boxes, which have notoriously lax security configurations. Anthropic believes that this attack method could become real “within months.”
The idea of AI-powered botnets may seem scary, but the report is more of a warning than a dire prediction. The results were obtained in a controlled environment, rather than on the open internet, and the predictions for when we’d see this in the wild are fairly vague. Additionally, the speed of such attacks is relatively slow, with times of 10 to 30 hours for spread, and 28-71 hours at minimum and around 51 to 150 hours maximum for getting the AI models fully operational on newly infected devices.
Protective Measures
Cybersecurity defenders have more time to quarantine infected devices, and bad actors have a disincentive to go faster, as spreading quicker makes detection easier. Furthermore, the use of connected devices with higher performing hardware appears to be more difficult, as their configurations aren’t as simple as with cheap internet products. Consumers can take steps to strengthen their device security, such as keeping devices off the internet if not in use, changing default passwords, and isolating devices on their own subnetwork when feasible.
Even if hackers start using small AI like this, consumers can still take protective measures. Keeping devices and software up to date and being cautious when clicking on links or downloading attachments can help prevent attacks.
