Are legions of smart devices – refrigerators, ovens, vacuum cleaners, lawn mowers, security cameras and televisions – conspiring to take over the Internet? No, but they may soon become unwitting hosts of self-replicating AI models that spread from device to device and prove remarkably difficult to contain.
Welcome to the troubling world of AI worms, where, depending on who you ask, the prospect of models spreading uncontrollably across household appliances is either a looming cybersecurity risk or another case of AI companies spreading catastrophic hype. In a blog post published this week, Anthropic warns that small AI models could take “months” to acquire such capabilities. His experiments offer some supporting evidence, but leave a huge gap between what the researchers demonstrated and the looming threat they predict. Yet the experts who spoke The edge say that while computer worms and fears of self-replicating AI are not new, recent progress suggests there are good reasons to take these warnings seriously.
Self-replicating machines have long haunted AI security researchers. They raise the prospect of the spread of systems beyond human control and difficult, if not impossible, to stop. As with many fears about advances in AI, attention has largely focused on increasingly powerful pioneering models, such as OpenAI’s Astra or Anthropic’s Mythos. But their size constitutes a major limiting factor: they are colossal. Containing potentially billions of parameters (values learned during training), boundary models require vast computing resources, confining them largely to data centers, where they can be more easily monitored and, in principle, deactivated.
Smaller models, however, present a different problem. They are improving rapidly, can run on consumer hardware, and are often available for free download. Anthropic researchers wanted to know if it could operate on a network of smaller, less powerful devices. They managed to split Google’s Gemma 4, a model with 26 billion settings, across 15 Android Smart TV boxes, none of which are powerful enough to run it alone. It worked, albeit very slowly.
After showing that a model could run on home devices, the researchers wanted to see if AI could perform the steps necessary for self-replication in such an environment. They built a simulated network of computers designed to behave like Android devices, then launched an AI-powered “worm” — a type of malware that automatically replicates and spreads across networks — tasked with finding new machines, copying the AI model to them, and running it on each one. It should be noted that the propagated model – Meta’s 8 billion parameter Llama 3.1-8B – was different from the one coordinating the attack, and much smaller.
Only the most sophisticated model tested, the GLM-5.2 from Chinese company Z.ai, managed to spread across the entire network, although smaller models managed to pass some milestones. Given how quickly they are improving, Anthropic researchers predict that smaller models could match the capabilities of the GLM-5.2 by the end of the year.
“As a prediction that goes beyond our experiments, we expect a future flooded with agents operating on all types of hardware, including some that are not under human control,” the researchers wrote. They acknowledged that they had not yet demonstrated “a completely malicious replicator,” nor did they know what it might look like (they suggested a “slow, distributed digital slime mold” or “a swarm of fruit flies”). But they said their findings showed “signs of life for AI worms” and called for stronger security on the many devices a replication system could use to spread.
“We should aim to harden our systems to make it difficult for these replicators to spread, but also prepare for a world where they are here to stay,” they wrote. It is important to note that Anthropic did not use its own models for these tests and has repeatedly argued that open alternatives to its AI models carry “more risk” than closed systems.
This warning sparked familiar accusations of AI hype and doom and gloom, particularly from readers of The edgethe front cover of . After all, there’s a huge gulf between showing that an AI can run slowly on a bunch of cheap TV boxes and a self-replicating AI worm tearing apart home devices. But while it’s certainly worth questioning a company’s motives, claims of deliberate scaremongering are difficult to substantiate, especially since Anthropic has long expressed concerns about the safety of AI. Nor is Anthropic alone in investigating the threat of self-replicating AI or calling for better security on connected devices.
Earlier this year, Palisade Research, a California-based AI security research firm, demonstrated that AI models could autonomously copy themselves to other machines. “What’s new here is that these models were much smaller, in terms of number of parameters, than previous ones, which were shown to be able to self-replicate,” said Jeffrey Ladish, executive director of Palisade. The edge. Anthropic’s work goes even further by showing that an AI can run on multiple low-power devices. “This will also be true for large models,” Ladish said, predicting that even the most advanced models could eventually run on large arrays of consumer chips, such as those found in MacBooks.
“Previous research has already demonstrated that the threat of AI-driven computer worms is realistic. »
Computer worms are not new either. They have been infecting machines for decades. But AI could make them smarter, more adaptable and better equipped to overcome any obstacles. Ladish noted that even smaller models are increasingly capable of carrying out cyberattacks autonomously. “Previous research had already demonstrated that the threat from AI-driven computer worms was realistic,” said Nicolas Papernot, a professor at the University of Toronto who has studied them. The edge.
Together, the capabilities demonstrated by Anthropic could significantly expand the range of devices that can be targeted in cyberattacks. “Devices that were not a target before due to attackers’ limited human resources and time can now become targets, if not for the data they contain, with the aim of using it as a launching pad to attack more valuable targets,” Papernot said. And there is no shortage of targets: the Internet is full of poorly secured smart devices.
Papernot warned against anthropomorphizing AI, but said the threat from AI-driven computer worms is “concrete” and “deserves immediate action,” especially when it comes to securing critical infrastructure like hospitals and power grids.
“Unless there is a massive effort to secure devices of all kinds, it is highly likely that AI worms will have fertile ground to colonize,” Ladish said. “I think we’ll see the first of these in the wild in the next 12 months.”
Gn bussni
