Anthropic said Monday that small AI models are “months away” from gaining the capabilities needed to replicate themselves across internet-connected household devices.
The company warned that AI models could potentially infiltrate devices such as smart ovens and Android TV boxes, use their computing resources to run themselves, and then spread to additional hardware without relying on expensive data centers.
Anthropic said such a system could become difficult to shut down because there would be no single centralized point to disable it.
No Fully Rogue Replicator Exists Yet
Anthropic acknowledged that researchers have not yet created a “fully rogue replicator.” However, AI agents are becoming increasingly capable of completing individual steps required for autonomous replication and passing their adaptations to future copies of themselves.
The research adds to wider concerns about advanced AI systems potentially carrying out cyberattacks or operating beyond human control.
In September, improvements in frontier AI models led to industry warnings about possible existential risks.
Executives and researchers at Anthropic and OpenAI have repeatedly raised concerns about increasingly capable AI systems and called for stronger safeguards and measures to control the pace of development.
Anthropic Tested AI Across 15 Android TV Boxes
As part of its research, Anthropic tested whether AI could replicate itself using inexpensive consumer hardware.
One experiment involved running Google’s Gemma4-26B-A4B model across a network of 15 low-cost off-brand Android TV boxes.
Anthropic noted that devices of this kind often have weak security, potentially making them easier for an AI system to infiltrate.
If a model is too large to fit on a single device, Anthropic said it could divide itself into smaller pieces, or “shards,” and operate across several devices at the same time.
Anthropic Warns About AI Worms and Swarms
Anthropic said a future AI worm may not rely on only one model or replication method.
A more realistic worst-case scenario could involve a swarm of different models running across different types of hardware and using several ways to replicate themselves.
The company said it remains uncertain what autonomous AI replication would ultimately look like.
One possibility is a slow, distributed network running quietly across insecure devices. Another could involve a powerful cyber model rapidly breaking into data centers, while another approach could rely on scams designed to obtain computing credits.
Anthropic Expects More Uncontrolled AI Agents
Anthropic said it expects a future in which AI agents run across many kinds of hardware, including some systems that are not under human control.
The company warned that the emergence of slower self-replicating agents could also serve as an early warning for more aggressive systems capable of breaching better-protected machines or stealing API keys.
Anthropic said the growing ability of agents to complete individual parts of the replication process suggests autonomous AI replication could become a more serious concern as models continue to improve.
The post Anthropic Warns AI Will Go Out of Control by End of 2026 appeared first on ProPakistani.