Anthropic warns AI models could soon copy themselves across home devices

Small artificial intelligence models are months away from developing the capabilities to replicate themselves across internet-connected devices in the home, Anthropic said Monday. The frontier lab detailed how AI could potentially infiltrate common smart devices, such as smart ovens or Android TV boxes, to utilize their local computing power. By spreading to other connected hardware, these models could sustain themselves without relying on expensive data centers. Anthropic noted in a blog post that such a system would be challenging to shut down because there is no single centralized point of control to turn off.

While the company acknowledged that a fully rogue replicator does not yet exist, it stated that current agents are becoming increasingly capable of completing the individual steps in the replication chain. These agents are also able to pass on their adaptations to future copies of themselves. This development adds to growing safety concerns regarding the ability of AI models to carry out cyberattacks and the potential for technology to operate without human oversight.

To test these capabilities, Anthropic conducted experiments using a small Google model named Gemma4-26B-A4B. The model was run across a network of 15 cheap, off-brand Android TV boxes. The company observed that such devices often possess weak security, creating vulnerabilities that a model could exploit. For models too large to fit on a single device, Anthropic explained that they could split into pieces, or shards, and run across multiple devices simultaneously.

The lab cautioned that a realistic worst-case scenario might not involve just one model or strategy. Instead, it could consist of a swarm of different models deployed on various types of hardware using diverse replication schemes. Anthropic stated it remains uncertain exactly what form autonomous AI replication will take. The company warned it could manifest as a slow, distributed system on insecure devices, a powerful cyber model that breaches data centers at a high pace, or a model that scams people for compute credits.

Anthropic predicted a future awash with agents running on all types of hardware, including some not under human control. The company added that the rise of slow replicators would serve as a bellwether for more aggressive, faster ones. These advanced systems could potentially breach more hardened machines or act as effective API key thieves. This research follows industry-wide warnings issued in September regarding existential risks to humanity stemming from improvements in frontier models.

Executives and researchers at both Anthropic and rival OpenAI have repeatedly warned about the risks of increasingly powerful AI. Some have called for tighter guardrails and measures to control the pace of technological development.