Until now, fears of AI copying itself from one machine to another across the internet have mostly been the stuff of thought experiments. Anthropic has now given that fear an approximate date. According to the company, if small AI models keep improving at today’s rate, they may be able to copy and spread themselves across connected devices by the end of 2026.
Anthropic's experiments did not aim at racks of servers in a data center. Instead, they targeted Android TV boxes and smart home devices, the kind of inexpensive gadgets that often sit on a shelf for years while nobody checks on them.
Small models handle individual steps, not the full chain
The research focused on one narrow question: could an AI system locate vulnerable devices, install copies of itself on them, and keep spreading with no human intervention whatsoever?
Models with fewer than 30 billion parameters produced mixed results. They managed to finish individual steps of the process, but they had trouble linking the whole sequence together.
Bigger systems performed better. GLM-5.2 and more powerful models carried out full end-to-end self-replication. However, this only happened in simplified tests where some security measures had been removed.
Check the caveats before you panic
That condition is important. If a model can only complete the task after some defenses have been stripped away, it has not shown that it can get into a properly locked-down network. What the findings do suggest is that the distance between “can do each step” and “can do all of them” is narrowing.
The late-2026 projection rests on a single assumption: small models must keep improving at their present pace. Anthropic chose the word “may,” and that word is doing much of the work here.
A slow spread that would be hard to halt
Even if self-replication does arrive, Anthropic is not predicting a rapid outbreak. The company expects any spread to happen slowly and to depend on widely available computing resources that have low security standards.
The harder problem is containment. Anthropic cautioned that without a central “kill switch,” these models would be difficult to contain. When a copy is running on a device you don’t control, there is no single place from which every copy can be shut down.
Anthropic’s proposed fix is not a new safety feature built into the models. The company is instead calling for stronger device security everywhere, which places much of the responsibility on hardware makers and on the people who own the devices.
The practical advice is simple. If there is an Android TV box or smart home hub on your network that hasn’t received a firmware update since you unboxed it, that is exactly the kind of hardware this research is talking about. Either update it or disconnect it from your network.