Anthropic warned self-replicating AI on home devices could arrive within months

Anthropic warned Monday that within months, compact AI models could grow powerful enough to spread autonomously across commercial internet-of-things devices — including smart ovens and off-brand Android TV boxes — without any human involvement.

The AI lab said models of 30 billion parameters and under can already complete many of the individual steps needed for replication, though they struggle to execute the full sequence end to end. In simplified tests with some security barriers removed, Anthropic found that a model called GLM-5.2 and stronger models were capable of end-to-end replication.

Halting such a system would present serious difficulties, Anthropic said, given that the architecture would lack any single point where it could be switched off. Unlike conventional malware, an AI-based replicator could also adapt to unfamiliar devices on its own.

Anthropic researchers William Zhou and Daniel Freeman said that if small models continue improving at their historical rate, they could match the capability threshold for autonomous replication by the end of this year.

In one demonstration, Anthropic deployed Google $GOOGL's Gemma4-26B-A4B model simultaneously on a cluster of 15 inexpensive off-brand Android TV boxes, each priced at around $55. Anthropic noted that consumer IoT hardware makes for appealing targets given the lax security standards that tend to govern such products. Where a model exceeds the capacity of any one device, Anthropic said it could be partitioned and its components executed in parallel across several machines.

The spread of such a worm would be slow. Anthropic estimated that a replicator based on Gemma4-26B-A4B would have a doubling time measured in days. The company compared it to a "digital slime mold," trickling tokens out across edge hardware rather than racing through networks.

Anthropic also tested whether AI replicators could adapt as they spread. In one experiment, successive generations of a Qwen3.5-9B model inherited notes from earlier instances, raising the inference setup success rate from 45% in the first generation to between 61% and 84% in later ones.

The researchers cautioned that the most dangerous outcome would not necessarily be a single unified worm, but rather a heterogeneous collection of models scattered across diverse hardware and employing varied methods to propagate. They also noted that future replicators could target data center processors, steal API credentials, or earn compute credits through freelance digital work.

Anthropic said the company does not yet have evidence of a fully functioning rogue replicator but that agents are becoming capable of completing individual steps in the replication chain and passing adaptations on to copies of themselves. The lab called for hardening IoT systems and said organizations should prepare for a world where some AI agents operate outside human control.