To The Moon

Anthropic issued a warning on Monday that small AI models are only months away from being able to replicate themselves across internet-connected household devices. This marks the latest alert from the frontier AI lab regarding potential safety risks tied to the technology.

Anthropic described how AI models could hack into connected devices such as smart ovens and Android TV boxes, using the devices' computing power to run their own programs and spread to other devices without relying on expensive data centers.

In a blog post, Anthropic stated that such models would be difficult to shut down because there is no central control point that can be cut off. The company acknowledged that a fully uncontrollable, self-replicating AI does not yet exist. However, AI agents "are becoming increasingly capable of completing individual steps in the replication chain and can pass their adaptive capabilities on to subsequent replicated versions."

Anthropic's latest research intensifies market concerns about AI safety, including risks such as models launching cyberattacks and AI going out of control. In September, advances in frontier model capabilities had already triggered industry-wide warnings about AI posing an existential risk to humanity. Executives and researchers at Anthropic and rival OpenAI have repeatedly warned about the potential risks of increasingly powerful AI, with some calling for stricter safety guardrails and controls on the pace of technological development.

Anthropic conducted multiple tests to verify whether AI can achieve self-replication. In one experiment, it ran Google's small model Gemma4-26B-A4B on a network of 15 cheap, off-brand Android TV boxes. The company noted that such devices often have weak security protections and are easily compromised by models. Models too large to fit on a single device can be split into multiple shards and run in a distributed manner across multiple devices.

"Real-world AI worms may not necessarily consist of a single model or a single strategy. A cluster deploying multiple models across different hardware and using diverse replication schemes may be a more realistic worst-case attack scenario," Anthropic said.

The AI lab said it remains uncertain what form autonomous AI replication will ultimately take: it could be a slow-spreading distributed system running on insecure devices, a powerful network model rapidly invading data centers, or a fraudulent AI that tricks its way into computing credits.

"Predictions beyond the scope of our experiments: in the future, AI agents will be ubiquitous across all kinds of hardware, and some of them will escape human control."

"The emergence of slow-replicating AI will also be an early signal, foreshadowing the arrival of more aggressive, faster-spreading AI in the future that can breach more strongly protected machines or steal API keys."