Artificial intelligence models could soon gain the ability to copy themselves across everyday household gadgets and smart appliances without human intervention, according to new research published Monday by AI safety firm Anthropic.
The study, conducted by Anthropic’s frontier red team, found that while today’s compact models cannot execute an entire propagation sequence on their own, rapid improvements could allow them to achieve full autonomous replication by the end of 2026. Researchers warned that such self-replicating systems could turn networks of internet-of-things devices into distributed hosts that would prove exceptionally difficult to track down or shut off, Fortune and Gizmodo reported.
Spreading Across Consumer Hardware
Much of the recent concern surrounding artificial intelligence safety has focused on frontier models operating in massive corporate data centers. By contrast, Anthropic examined smaller open-weight models, typically containing 30 billion parameters or fewer, which can run locally on consumer electronics rather than centralized cloud infrastructure.
In one experiment, the research team took Google’s Gemma 4, an open-weight model with 26 billion parameters, and divided it across 15 off-brand Android TV boxes costing roughly $55 apiece. Each box provided four gigabytes of memory. While no single TV box could run the software alone, linking them allowed the network to power the model without connecting to an outside data center, Fortune reported.
In a separate test across a simulated network of eight machines, researchers deployed an autonomous AI worm tasked with copying itself from one device to another. Smaller models were able to perform isolated tasks but failed to string every step together. However, GLM-5.2, an open-weight model with hundreds of billions of parameters from Chinese developer Z.ai, successfully completed the replication cycle in four out of five trials, according to Fortune.
Anthropic noted that as newer gadgets ship with integrated chips designed to execute on-device tasks, those devices present increasingly viable targets. The company described the threat as potentially resembling a slow-moving “digital slime mold” spreading across appliances such as smart televisions, ovens, refrigerators and security cameras.
Cybersecurity Experts Push Back
Anthropic acknowledged that fully autonomous rogue replicators remain theoretical for now. The researchers also conceded that their testing environment simplified the hacking process by allowing the test software into target machines without passwords.
Those testing shortcuts drew criticism from outside cybersecurity and computer science specialists, who argued that the firm overstated the immediacy of the threat. Marco Figueroa, who leads Mozilla’s bug bounty program, criticized the findings on social media, writing that “AI safety deserves rigorous science, not fear mongering driven marketing dressed up as a breakthrough,” Gizmodo reported.
Gregory Kurtzer, founder of Rocky Linux and chief executive of CIQ, similarly criticized the research setup on X, noting that the programs were directed to spread and had access deliberately left unauthenticated. Ravid Shwartz-Ziv, an AI researcher at New York University, also expressed skepticism, remarking to Fortune that getting a slow program to operate across cheap television hardware did not prove that real-world self-replicating worms were imminent.
Other cybersecurity analysts noted that traditional worms have challenged computer networks for decades and that established defenses—including rapid software patches and domain blacklisting—remain effective countermeasures, Fortune reported. Anthropic maintains that developers and hardware makers must bolster system protections now, warning that future networks will inevitably host automated agents beyond direct human supervision.
