AI models could learn to spread through smart home devices

The company Anthropic has warned that within the coming months, small artificial intelligence models could learn to copy themselves on their own and spread through devices connected to the internet.

Such a scenario could significantly complicate efforts to counter cyberattacks, since a malicious system could potentially establish itself on a large number of devices that are not centrally controlled.

This concerns Internet of Things devices, which include, in particular, televisions, household appliances, and other network-connected gadgets. According to the company's specialists, advances in artificial intelligence could allow small models to carry out the operations needed for independent propagation in the foreseeable future.

Today, most powerful artificial intelligence models run on servers and in data centers managed by their developers. If a serious threat arises, a company can shut down the system in question. However, the situation would become considerably more complicated if a model were able to copy itself onto thousands of devices in homes and offices. In that case, stopping its operation would require finding all infected devices and removing the corresponding software from them.

Anthropic notes that it is still impossible to say with certainty what autonomous self-replication of artificial intelligence would look like in practice. At the same time, the assessments conducted show that models with up to 30 billion parameters are already capable of performing many of the individual actions required for copying and spreading. However, when carrying out the entire chain of operations in sequence, they still make mistakes.

If the pace of development of small models continues, they could learn to perform such tasks reliably by the end of 2026.

According to Anthropic's specialists, self-spreading artificial intelligence could pose a threat not only because of the possibility of establishing itself on a large number of devices. Such systems could potentially automate the search for vulnerabilities, accelerate the hacking of computer infrastructure, and gain access to computing resources without the permission of their owners.