Anthropic warns AI could aid malware spread, though no model has fully self-replicated
Anthropic says smaller AI models could gain the ability to copy themselves between connected devices within months, raising concerns about malware that may be difficult to contain. The company’s experiments offer some evidence that AI can help with parts of this process, but researchers have not demonstrated a fully self-replicating system.
The researchers split Google’s 26-billion-parameter Gemma 4 across 15 Android TV boxes, where it ran slowly, then tested AI-powered malware in a simulated network. Only the most capable model they tried, GLM-5.2 from China’s Z.ai, spread across the entire network; other models managed individual steps. Anthropic predicts smaller models could match that capability by year’s end, while experts say the risk merits attention despite the gap between the tests and the forecast.
- Anthropic warns small AI models could spread between devices.
- Tests showed partial capabilities, not a full replicator.
- Experts say device security deserves attention.
New here? Start with this
AI models are computer systems trained to produce text or carry out tasks. Malware is software designed to disrupt devices or gain access without permission. A self-replicating program can make copies of itself and spread to other connected devices.
Anthropic is an AI company that has tested whether models can help malware perform steps involved in spreading. Its work involved a model distributed across Android TV boxes and a separate test in a simulated network. These are experiments, and researchers have not shown an AI system that can independently copy and spread itself from device to device.
The issue matters because connected devices can give malicious software routes to reach more systems. Security researchers and technology companies study such risks to understand what safeguards may be needed, while distinguishing demonstrated capabilities from predictions about what models may do later.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The experiments show that AI can already help with parts of malware propagation, and one tested model spread through the simulated network. As smaller models become more capable, defenders may have little time to prepare; early attention and investment in safeguards could limit harm before a fully self-replicating system appears.
The case against
The tests did not demonstrate a fully self-replicating system, and only the most capable model spread across the simulated network, so the forecast that smaller models will soon do so remains uncertain. Treating that prediction as imminent could divert attention and resources from threats that are already established, while encouraging measured monitoring and evidence-led preparation.
Read the full article at the source →
Originally published by The Verge as “Anthropic’s self-replicating AI warning is an alarming deja vu”.