Anthropic recently published a disturbing research report on its official website: AI models may, in just a few more months, gain the ability to self-replicate, and once they spread, they will be very hard to stamp out completely. The reason is straightforward—such programs have no centralized node that can be shut down, and they scatter everywhere like weeds.
What makes them even trickier is where they settle. Experiments show that a model with only 90 hundred million (9 billion) parameters can already run inference on 8 of 9 tested hardware classes, handling even commercial IoT devices like smart ovens and cheap Android TV boxes with ease. Because these devices are cheap and thinly protected, they are becoming a breeding ground for "digital slime mold"-style parasites. Anthropic estimates that models with no more than 300 hundred million (30 billion) parameters can currently perform many of the individual steps of self-replication, but often make mistakes when stringing them together end-to-end; in simplified evaluations, however, GLM-5.2 and stronger models have already completed the full chain. Extrapolating from past rates of progress, small models could reach this threshold by the end of this year, after which replicators based on small weights could quickly become reality.
When ARA becomes real, the industry will have to learn to live with the risk
This phenomenon, known as "Autonomous Replication and Adaptation (ARA)," works much like worms in cybersecurity: without human intervention, it scans networks, exploits vulnerabilities, and copies model weights onto target machines. The difference is that AI not only carries its own weights but can also learn on the fly, figuring out how to run on unfamiliar devices. Data center GPUs are certainly juicy targets, but distributed IoT botnets can likewise sustain surprisingly capable models.
Anthropic stresses that no truly out-of-control, fully self-replicating programs have yet been found, but agents are getting better and better at this, and they pass adaptive improvements on to the next generation of copies. The company predicts the future will be filled with all kinds of uncontrolled agents—slowly parasitic "digital slime mold," aggressive models attacking data centers, or even variants that deceive humans into handing over compute. One warning sign worth heeding: a steady increase in slow-replicating programs often presages the arrival of faster, more aggressive versions. Anthropic recommends that the industry harden systems and do its best to block propagation, but also prepare for a world in which "they will exist for the long term."