Researchers at Anthropic have warned that future artificial intelligence systems could potentially gain the ability to replicate themselves across internet connected devices without direct human control. The findings, published by Anthropic’s Frontier Red Team, examine the possibility of self replicating AI systems that could discover devices, transfer model files, install required software, and continue operating from newly accessed machines. While researchers emphasized that current AI models do not represent a fully autonomous replication threat, the study highlights how rapidly advancing AI capabilities could introduce new challenges for cybersecurity and digital infrastructure. The research explored whether smaller AI models could independently complete the technical steps required for replication, particularly across devices with limited security protections.
According to Anthropic’s research, AI models with fewer than 30 billion parameters are already capable of completing several individual stages involved in a potential self replication process, although most systems still struggle to combine those steps consistently. In a controlled and simplified experiment, the GLM 5.2 model successfully spread across all eight available test devices in four out of five trials, while reaching seven devices in the remaining trial. Anthropic clarified that the experiment included simplified conditions, including devices with unsecured connections that reduced the complexity of the hacking process. The company noted that these results do not demonstrate that an AI model can currently escape into real world environments and reproduce itself independently. Researchers described the current situation as an early stage where AI agents are becoming more capable of performing tasks required for replication and sharing useful information with future versions of themselves.
The study focused on inexpensive Internet of Things devices, which researchers identified as potential targets because many consumer devices often have limited security protections. These devices could include smart televisions, refrigerators, ovens, robotic vacuum cleaners, security cameras, and video conferencing equipment. Anthropic demonstrated that a 26 billion parameter model could operate across 15 low cost Android television boxes, each costing around $55. The research also showed that a smaller model was able to establish the software environment needed to function across eight of the nine hardware types tested. Unlike traditional computer worms that typically rely on fixed methods of spreading, researchers said a self replicating AI system could potentially adapt to unfamiliar devices, identify technical challenges, and transfer successful strategies to later copies, making such systems more difficult to contain.
Researchers explained that an early AI replicator would likely spread slowly, potentially taking one or two days to create each new copy. However, they warned that eliminating such a system could become challenging because it may not depend on a central server that can simply be disabled. Anthropic estimated that if smaller AI models continue improving at their current pace, they could reach the capability demonstrated by GLM 5.2 by the end of 2026, although the company stressed that this projection is based only on general capability trends and should not be considered a prediction. To prepare for potential future risks, researchers recommended monitoring unusual network activity, isolating affected devices, and encouraging manufacturers to improve default security protections. They added that early, slow moving AI replication systems could provide an indication of future technologies capable of interacting with more secure computing environments. Anthropic said organizations should focus on strengthening systems to reduce opportunities for such replicators to spread while also preparing for a future where increasingly capable AI systems become part of the broader technology landscape.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
