Anthropic: Small AI Models Could Create Networks of Infected Devices

    According to Anthropic’s Frontier Red Team, small artificial intelligence models may be only a few months away from spreading by creating copies of themselves on poorly secured devices. Such models could create autonomous botnets that spread slowly but are difficult to disable. A botnet is a remotely controlled network made up of infected devices. Experts say the threat is real, but its consequences are still unclear.    Chris Canal, head of EquiStamp, expects this stage to be reached by next summer. In his view, Anthropic may have more detailed internal information about the current state of the research. One factor accelerating development is that AI labs use their own tools to develop new models. The company’s formal filing for an IPO, or stock market listing, after the research has raised questions about whether it was intended for marketing purposes.    The threat is not entirely different from previous botnets. In 2016, Mirai infected internet-connected devices such as routers and IP cameras and carried out a DDoS attack on the PlayStation Network. During such an attack, an excessive number of requests are sent to a server, causing the service to stop working. The incident also affected Dyn, a major DNS service, and took numerous services offline. Sony’s net income loss was $2.7 million (approximately 4.6 million manats).    The main difference with an AI-based botnet is its ability to adapt to unfamiliar devices. Although Mirai ran on various processors, the malware had to be rewritten by a person for each new system. An AI model could automate this process. However, the test was conducted on a virtual network with a simple security vulnerability. It was also not demonstrated that a complete copy of the model was transferred to every device.    During the experiment, the model was divided into parts and run on separate devices. Its copies then passed notes to models launched later. For this reason, the researchers compare the system not to a complete program that spreads by itself across the internet, but to a small management tool that installs other code on devices. The model would have had to retrain itself to change its footprint and behavior, but no such test was conducted. It is also currently difficult to impose reliable restrictions on models with open weights.    Anthropic is urging internet service providers to notify infected customers and restrict their access to secure websites. In the United Kingdom, easily discoverable default passwords on internet-connected devices are already prohibited by law. Experts cite timely device updates, separating networks from one another, and using strong passwords as the main methods of protection. AI may make attacks easier, but spreading to protected devices will remain difficult.