MUMBAI: The next computer virus may not need a computer room to thrive, it could find a home in your living room. Anthropic has warned that smaller artificial intelligence models could potentially spread across internet-connected household devices within months, raising concerns about how to contain AI systems that become capable of replicating themselves.
In research published on Monday, the AI company explored how models could potentially harness the computing power of devices such as smart ovens and Android TV boxes to operate across multiple machines. Such a development could allow AI systems to function beyond conventional data centres, making them harder to monitor, isolate or switch off.
There is an important caveat, however: Anthropic said a fully autonomous AI system capable of independently replicating itself has not yet emerged. Its warning centres on the capabilities AI agents are developing and the potential risks if those abilities are combined.
The company said AI agents were becoming increasingly proficient at individual stages of the replication process, including transferring modifications to subsequent copies. If these capabilities were combined into a functioning system, it could create a network without a single point of control, complicating efforts to shut it down.
To examine the potential vulnerabilities, Anthropic conducted experiments using Google’s Gemma 4-26B-A4B model across 15 low-cost Android TV boxes. The research highlighted how inadequate security on connected devices could make them potential targets for infiltration.
Anthropic also explored how models that exceed the computing capacity of a single device could be split into smaller components. Different machines could then contribute computing resources to support the wider system, rather than requiring one device to handle the entire workload.
That opens up a more complex cybersecurity scenario. Instead of a single programme following one replication strategy, an attack could involve multiple models running on different hardware and using different methods to spread.
“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 potential risks extend beyond household gadgets. Anthropic outlined scenarios ranging from distributed networks operating on vulnerable devices to more capable AI systems compromising data centres or obtaining computing resources through fraudulent means.
The company warned that slower, less conspicuous replication could also signal the possibility of faster and more aggressive systems capable of breaching better-protected machines or stealing API keys — credentials that allow software to access online services.
“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,” Anthropic said.
The eventual form of autonomous AI replication remains uncertain. The scenarios described in the research are potential risks, not evidence that a self-replicating AI network is already operating.
The warning adds another dimension to the debate over increasingly capable AI systems. As AI agents become better at carrying out multi-step tasks, questions about their ability to access resources, operate across systems and evade containment are becoming more pressing.
Anthropic and rival OpenAI have repeatedly raised concerns about the risks associated with more powerful AI models, fuelling calls for stronger safeguards and careful management of AI development.
For now, the prospect of an AI worm hopping from one household device to another remains a potential scenario rather than an established threat. But Anthropic’s research points to a broader challenge: as connected devices multiply, securing the AI systems that might use them could become just as important as securing the networks they run on.