AI 'worms' may self-replicate in homes!

Attention has typically focused on rogue AI agents targeting corporate websites and systems, but smaller targets, such as the smart home network, may soon come under fire. This is the warning issued by futurism.com in a report based on new research from Anthropic.

Anthropic stated in a report published on Monday that “AI worms” could soon be capable of breaching televisions, cameras, and other devices connected to wireless networks, using them as incubators to replicate and spread to other devices. The company noted that this threat may be only months away if models continue to improve at their current pace. Futurism pointed out that the company is no stranger to issuing grim and dramatic warnings.

The report stated, “We anticipate a future saturated with agents operating on all types of hardware, including devices not under human control.” Worms are malicious software that replicate themselves and spread to other devices by searching for those connected to the same network. Anthropic researchers are specifically interested in how rogue agents, which may soon acquire this capability, spread across everyday hardware.

Many homes appear poised for targeting as the proliferation of internet of things (IoT) devices connected to home networks expands, including smart TVs, refrigerators, lighting, robotic vacuums, surveillance cameras, and even toilets. According to Futurism, the vulnerability of these systems stems from several factors: devices manufactured by companies whose reliability is questionable, users who may lack sufficient technical awareness, networks that may not have been secured properly, and processing chips that are becoming more powerful even in the simplest consumer devices.

If these factors converge, a small AI model could infiltrate these devices and run itself on them. The website noted that previous research found that a potential AI worm could carry its model weights during replication, then begin running them on the hardware it takes over.

- A lightweight model successfully operated across 15 smart TV boxes, each with no more than 4 gigabytes of random-access memory.

- AI-based worms can understand unfamiliar devices without any assistance, unlike non-AI malware.

- Replication of these models on lightweight home devices is expected to be slow, making them more akin to “digital slime mold.”

The report added that graphics processing units (GPUs) in data centers would be a “clear and high-value target” for such malware, but IoT devices, which are typically cheap and poorly protected, could provide sufficient computing power to run an efficient model.

Researchers warned that although IoT devices have long been vulnerable to hacking, lightweight agents may be far more willing than human attackers to expend the effort required to compromise a significant portion of the computing power of these devices worldwide.

The report concluded that the picture is not yet complete, as the world still lacks a “fully autonomous replicator,” but agents are becoming increasingly capable of executing each step of the replication chain individually.