AI Set to Invade Homes? Anthropic's Worrying Scenario for Coming Months
Artificial intelligence may soon no longer need massive data centers to replicate itself, instead using the processing power of everyday devices to spread. Researchers at Anthropic have warned that small AI models could be capable of replicating themselves by the end of 2026 on devices such as smart TVs, refrigerators and other internet-connected equipment, a development that could make cyberattacks harder to contain.
technology group; Researchers at Anthropic, maker of the Claude family of artificial intelligence models, have warned that small artificial intelligence models may become powerful enough to be deployed on Internet of Things (IoT) devices in homes and small businesses within the next few months. Without needing large data centers.
According to the report published by the company's researchers on Monday, the development of small models of artificial intelligence can enable the implementation of such a scenario by the end of 2026. This research has relied on the testing of several artificial intelligence models developed in America and China.
However, the researchers emphasized that in the tests conducted, none of the current models were simultaneously powerful and compact enough to reliably replicate on the smart TV being tested. Therefore, what has been proposed is a prediction about the way technology will progress, not a confirmation that such an attack will happen now.
Why TV and refrigerator can become targets of attacks?
IoT devices include a wide range of equipment connected to the Internet; From smart TVs, refrigerators and ovens to lawnmowers and small business equipment. These devices are typically designed to perform specific tasks and, in many cases, are not given the same level of cybersecurity attention as enterprise computers and servers. This issue can make them attractive targets for cyber attacks.
William Zhou and Daniel Freeman, Anthropic researchers, believe that IoT devices in homes and small businesses may be among the first targets of such attacks; Because they are connected to the Internet and in the past they often had limited security measures.
If this scenario is realized, potential attackers can use the security vulnerabilities of these devices to run artificial intelligence agents. Such agents may try to identify and infiltrate other devices in the next step, and gradually a wide network of infected equipment is created.
What is an artificial intelligence worm and why is it difficult to contain it?
One of the main concerns of researchers is the emergence of a type of malware called "AI Worm"; Malware that can use the reasoning capabilities and adaptability of artificial intelligence models to expand itself.
Computer worms have a long history and were able to spread between devices through networks before the advent of artificial intelligence. The possible difference of the new generation of these malwares is that artificial intelligence agents can be more capable in identifying unknown environments, changing methods after failure, and adapting to new conditions.
Roman V. Yampolsky, an artificial intelligence and cybersecurity expert at the University of Louisville, explained to Al Jazeera that an AI worm might be able to do something beyond traditional malware: reason about unknown devices, change its method when it fails, and continue to spread without direct human guidance.
According to him, such a threat is not an existential threat to humanity in the current situation, but the path of its development is worrying.
Yampolsky warns that if automated replication is combined with hacking capabilities, persistence in infected systems, and strategic reasoning, it will become more difficult to stop; Especially if there is no longer a single server or central organization that can shut down the entire network.
Why the absence of a control center can be dangerous?
An important part of Entropic's concern relates to the difference between a centralized infrastructure and a network of distributed devices. Large data centers, despite their complexity and strategic importance, are specific targets whose activity can be identified and limited in certain circumstances. But if AI agents can run on a large number of dispersed home and business devices, it will be more difficult to identify and stop them all.
Anthropic researchers believe that data centers will be obvious and valuable targets for artificial intelligence worms, but Internet-connected devices may also become alternative processing resources for the proliferation of these agents.
This company has predicted that the future can witness the widespread presence of artificial intelligence agents on all kinds of hardware; including devices that are not necessarily under direct human control.
However, a distinction must be made between running an AI agent on a device and its ability to independently replicate, infiltrate other devices, and continue to operate without human control. Realization of the second scenario requires certain technical capabilities and security conditions and is not yet a proven capability in home devices.
Warning about the spread of artificial intelligence without human control
Entropic's new report has been published amid growing concerns about the unwanted behavior of artificial intelligence models and their use in cyber attacks. Anthropic and its main competitor, OpenAI, have previously called for the international community to coordinate to slow down the development of artificial intelligence in order to limit catastrophic risks. However, this proposal has not been widely welcomed by the US and China, the two leading powers in this field.
Some defenders of artificial intelligence safety also mention "Recursive Self-Improvement" as one of the possible turning points of this technology; A situation where models can help improve their abilities and this process continues sequentially.
Recursive self-improvement is not the same as self-replication. Self-replication refers to the creation of other versions of a system, while feedback self-improvement refers to the process by which a system improves its own capabilities. The possible combination of these capabilities with advanced cyber capabilities is one of the concerns raised in the AI safety debate.
Last month, Anthropic and several major US technology companies entered into a non-binding agreement that emphasized strengthening internal safeguards and cooperation with outside evaluators. This agreement was met with different reactions and a number of experts criticized the voluntary nature of its obligations.
Entropic's claim about 90-99% reduction in unauthorized data mining
In a separate announcement on Monday, Entropic announced that it has made significant progress in preventing so-called "distillation attacks". In this type of attack, the outputs or capabilities of an advanced AI model are used to train cheaper and less powerful models, without proper authorization.
The San Francisco-based company has previously accused its Chinese rivals, including DeepSeek, Moonshot and MiniMax, of industrial-scale attempts to copy the cloud.
Antropik has announced that the measures it has taken to deal with these attacks have reduced the amount of unauthorized data extraction between 90 and 99 percent.