Anthropic Warns AI Models Could Self-Replicate Across Home Devices

Small artificial intelligence models could gain the ability to replicate across internet-connected household devices within months, according to research published Monday, October 12, by Anthropic. The San Francisco-based AI company warned that vulnerable smart appliances and television boxes could provide computing power for autonomous AI systems to spread without relying on centralized data centers.

Anthropic said such a network could become difficult to disable because it would lack a single central point of control. However, the company acknowledged that researchers have not yet demonstrated a fully autonomous AI system capable of reliably replicating itself across consumer devices.

The company described the prospect as a developing cybersecurity concern, noting that AI agents are increasingly capable of completing individual steps required for replication and transferring adaptations to subsequent copies.

Anthropic's research included experiments with Google's Gemma4-26B-A4B model running across 15 inexpensive, off-brand Android TV boxes. The company also described how models too large for individual devices could be divided into smaller components and distributed across multiple machines.

Anthropic cautioned that a potential attack might involve multiple models operating on different hardware rather than one system spreading independently.

"In practice, an AI worm wouldn't necessarily consist of just one model or strategy. A swarm of different models deployed on different types of hardware using a variety of replication schemes might represent a more realistic 'worst case scenario' attack," Anthropic said.

The company said the precise form of autonomous replication remains uncertain. Potential scenarios include a slow network of compromised devices, a powerful cyber model breaching data centers, or an AI system obtaining computing resources through scams.

"As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren't under human control," Anthropic said.

Anthropic also warned that slower replication could signal the emergence of more aggressive systems capable of penetrating better-protected machines or stealing API keys.

"The rise of slow replicators would also be a bellwether for more aggressive, faster ones that can breach more hardened machines or be effective API key thieves," the company said.

Researchers William Zhou and Daniel Freeman identified household and small-business devices as potential early targets because of their internet connectivity and historically limited security protections.

"Home and small business IoT devices might be the first target for such attacks because they're connected to the internet and, historically, have had minimal security measures in place," Zhou and Freeman said.

They argued that autonomous agents could devote substantial effort to compromising connected devices on a scale that would be difficult for human attackers to match.

"While it's true that IoT devices have been vulnerable for a long time, lightweight agents may be much more willing than human attackers to put in the effort to compromise large fractions of the world's IoT compute," the researchers said.

Roman V Yampolskiy, an AI and cybersecurity expert at the University of Louisville, said AI-driven malware could adapt its behavior to unfamiliar systems and continue operating without direct human involvement.

"Computer worms are not new, but an AI worm could do something traditional malware cannot: reason about unfamiliar devices, adapt when its approach fails, and continue spreading without a human operator," Yampolskiy told Al Jazeera.

He distinguished the current situation from a direct existential threat while warning that future advances could increase the danger.

"Today, this is not an existential threat. But the trajectory is concerning," Yampolskiy said.

Yampolskiy said combining autonomous replication with advanced hacking capabilities and persistent operation could make containment substantially more difficult.

"Once autonomous replication is combined with strong hacking, persistence, and strategic reasoning, shutting such a system down becomes much harder because there may be no single server or organisation controlling it," he said.

Anthropic's findings add to wider debate about the security implications of increasingly capable AI systems, including concerns about unauthorized cyber operations and the possibility of technology acting beyond human control.

Anthropic and OpenAI have called for international coordination to slow AI development to reduce catastrophic risks, although the proposal has received limited support from the United States and China, according to Al Jazeera.

Last month, major US technology companies, including Anthropic, joined a nonbinding agreement to strengthen internal safeguards and engage external auditors. The voluntary nature of the commitment drew criticism from some experts.

In a separate announcement Monday, Anthropic said it had reduced unauthorized data extraction associated with distillation attacks by 90% to 99% through measures designed to prevent competitors from using advanced AI systems to train cheaper models.