Anthropic Warns Of Self-Replicating AI As Google Seeks Safeguards And Cerebras Pushes Back

Artificial Intelligence (AI) leaders are painting very different pictures of what comes next.

Anthropic warns that smaller models could spread across connected devices without data centers, Google DeepMind wants stronger safeguards, and Cerebras says the industry is overlooking AI's benefits.
 

TL;DR

 

• Anthropic believes small AI models may become capable of self-replication by late 2026, although no uncontrolled outbreak was demonstrated.
• Google DeepMind calls for flexible global safety standards while Google Cloud expands research access in Europe.
• Cerebras CEO Andrew Feldman challenges catastrophic AI predictions and envisions personal AI assistants running partly at home.
• Outlook: Security, oversight, and tangible benefits will determine public trust.

 

Anthropic Warns AI Models Could Self-Replicate Across Smart Devices

Imagine an AI worm operating through inexpensive Android TV boxes rather than a powerful data center. That is the possibility explored in Anthropic's October 12 research, which argues that connected household devices could collectively provide enough processing power for autonomous AI.

Researchers ran a 26-billion-parameter model across 15 TV boxes costing about $55 each. In another evaluation, a 9-billion-parameter model set up inference on eight of nine hardware types.

Importantly, this was not a rogue AI spreading freely across the public internet. The replication trials took place in a controlled network with simplified access, where a larger model propagated a smaller model's payload across eight test nodes in four of five runs.

Anthropic says models with 30 billion parameters or fewer can already execute many individual replication steps, but cannot reliably combine them. If improvements continue, it estimates small models could cross that threshold by the end of 2026.

A distributed AI botnet could be harder to switch off because no single server controls it. Yet slow operation on inexpensive hardware and distinctive network traffic may also create detection opportunities.

Expert reactions have been mixed: former UK cybersecurity chief Ciaran Martin questioned the urgency, while University of Colorado Denver researcher Zhengxiong Li warned that adaptive AI worms could be difficult to shut down. The concern adds context to Anthropic's tightened Claude cyber evaluations.
 

Google DeepMind Urges AI Safety Standards While Google Cloud Expands Scientific Access

The debate also reached Dublin, where Google DeepMind co-founder Demis Hassabis described recent AI-driven hacks as “early warning shots” on October 14.
 

 

Hassabis warned that more serious incidents could follow unless companies and governments strengthen defenses. He called for international safety standards inspired by safeguards in fields such as biology and nuclear technology, but stressed that certification should adapt to changing risks rather than restrict innovation unnecessarily.

On the same day, Google DeepMind and Google Cloud announced plans to provide up to 200 European scientists access to frontier tools and $3 million in compute credits.

The initiative includes AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve, and AlphaEarth Foundations, supporting work on molecular structures and extreme-weather prediction. This is targeted scientific access, not a general release of frontier models.

Together, those positions reflect Google's pursuit of research benefits with stronger oversight, especially considering DeepMind Institute.
 

Cerebras CEO Challenges AI Doomsday Talk And Envisions Personal AI Assistants

Cerebras CEO Andrew Feldman offered a different emphasis at TechCrunch Disrupt 2026, arguing that relentless catastrophe predictions overshadow AI's possible benefits.

During the discussion, Feldman joked that “some of the leaders in the AI world maybe shouldn’t be allowed to talk in public.” Yet he acknowledged the risks of powerful technology, supported reasonable regulation, and criticized poor community engagement over data-center construction.

His optimism extends beyond industrial computing. Feldman also predicted that AI assistants will become home appliances, splitting their work between local hardware and the cloud.

These systems could learn users' preferences, access payment information, and buy things on their behalf. Feldman already uses agents for calendar management and meeting preparation, though a mass-market Cerebras home device was not announced.

That vision adds a consumer dimension to the security conversation: assistants become more useful as they gain permissions, but those permissions also increase the importance of safeguards.
 

 

Outlook: AI Security And User Trust Will Shape Wider Adoption

Anthropic's self-replication scenario remains an experimental warning, Google's proposed international safety framework remains a proposal, and Cerebras' household AI vision remains a prediction.

What unites them is the need to prove that more autonomous systems can be both useful and governable before they are trusted with connected devices, sensitive information, and purchasing decisions.