Anthropic warns that AI models could self-replicate on home devices

Anthropic said Monday that small AI models would be “months” away from developing the capabilities needed to replicate themselves on internet-connected devices at home, marking the latest warning from the border lab about the technology’s potential security risks.

Anthropic described how AI models could potentially infiltrate internet-connected devices, such as smart ovens or Android TV boxes, and use their computing power to operate on their own and spread to other devices, without relying on expensive data centers.

The AI ​​model would be “difficult to stop,” Anthropic said in a blog post, “since there is no centralized point to stop.”

The company acknowledged that there is no “totally malicious replicator” yet.

Agents are, however, “increasingly capable of completing different steps in the replication chain (and passing on their adaptations to future copies of themselves),” Anthropic said.

Anthropic’s latest research adds to growing security concerns around AI, such as the models’ ability to carry out cyberattacks and the technology becoming malicious. In September, improvements to the pioneering models sparked industry-wide warnings about existential risks to humanity.

Executives and researchers at Anthropic and rival OpenAI have repeatedly warned of the potential risks of increasingly powerful AI, with some calling for stricter guardrails and measures to control the pace of the technology’s development.

Anthropic said it conducted various tests to see if the AI ​​could self-replicate. One of these tests involved running a small Google model, Gemma4-26B-A4B, on a network of 15 off-brand cheap Android TV boxes.

The company noted that these types of devices often have weak security, which could allow a model to infiltrate them. A model too large to fit on a single device could “split into pieces or fragments” and work on multiple devices, Anthropic said.

“In practice, an AI worm would not necessarily consist of a single model or strategy. A swarm of different models deployed on different types of hardware using a variety of replication schemes could represent a more realistic ‘worst-case scenario’ attack,” Anthropic said.

It remains unclear exactly what form AI autonomous replication will take,” the AI ​​lab said, warning that it could be a “slow and distributed” system sitting on insecure devices or a “powerful cyber model that breaches data centers at a high rate, or a model that scams people for computing credits.”

“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,” Anthropic said.

“The rise of slow replicators would also be an indicator of the presence of more aggressive and faster replicators capable of hacking more hardened machines or effectively stealing API keys,” the company added.