Anthropic released research warning that compact AI models are edging closer to being able to autonomously self-replicate without relying on centralized data centers. Experiments revealed that models with 30 billion parameters or fewer can already handle many of the individual tasks required for autonomous propagation. The proposed threat relies on exploiting decentralized, lower-powered, and often poorly secured internet-connected devices like smart appliances and Android TV boxes. Because these self-replicating systems would operate across distributed networks instead of a central cloud server, defenders would lack a single point to shut them down. Anthropic cautioned that if small models maintain their historical pace of capability improvements, they could reach full autonomous replication thresholds by late 2026.

Full report : AI models may soon self-replicate without access to onerous compute warns Anthropic.