Anthropic Discovers a Dangerous AI Capability - Root-Nation.com
Smart Ovens and TVs at Risk: AI Models Could Self-Replicate as Early as This Year
© ROOT-NATION.com - Use of content is permitted with a backlink.
Anthropic says the results of its experiments indicate that small AI models capable of independently spreading copies of themselves through commercial Internet of Things (IoT) devices could emerge within just a few months. Potential targets include smart ovens, Android TV boxes from lesser-known manufacturers, and other internet-connected gadgets.
As part of the study, Anthropic tested whether AI models could identify vulnerable devices, transfer copies of themselves to them, and continue spreading without human involvement. The experiments showed that even models with fewer than 30 billion parameters can perform individual stages required for such a process.
At the same time, the researchers found that small models still struggle to complete the entire sequence of actions from start to finish. In simplified tests, where some safety checks were removed, GLM-5.2 and more powerful models demonstrated the ability to reproduce themselves autonomously at every stage.
Anthropic also warned that stopping AI systems capable of autonomously spreading copies of themselves could be difficult because they would not have a single central shutdown mechanism. Unlike conventional malware, such systems could independently figure out how to interact with unfamiliar devices. They could also potentially target the processors used in data centers.
In its study, Anthropic focused on Internet of Things devices, including Android TVs, smart home appliances, and other internet-connected gadgets. These devices were selected because their security is often inadequate, creating favorable conditions for the potential spread of AI models.
The company also warned that if small models continue improving at the same pace as before, they could gain the ability to self-replicate autonomously by the end of this year. Similar capabilities could emerge soon afterward in systems that use their own model weights for replication.
At the same time, Anthropic believes that such autonomous AI systems would spread slowly and rely on publicly available computing resources with relatively weak security. The company emphasized the need to strengthen system security to make it more difficult for such AI agents to spread between devices.
Read also: Meta Introduces Its Most Powerful AI Models, Muse Wonder: What They Can Do and Who Can Access Them