Anthropic warns of the risk of AI capable of replicating like computer worms2026-10-14Anthropic warns that small AI models could spread autonomously, like computer worms, infecting connected devices.In tests on around 15 low-cost smart TVs, researchers completed every step of replication, but the models failed to complete the process on their own.After the tests were simplified and some safeguards removed, a more powerful model was able to replicate fully.The risk remains hypothetical; researchers estimate that, if progress continues at its current pace, the simplest models could replicate within a few months.De petites IA qui se comportent comme des vers informatiques et se répliquent sans intervention humaine... (BFMTV)Anthropic advierte sobre la propagación de “gusanos” de IA en dispositivos inteligentes (Página/12)AI models can now hack gadgets independently, making human hackers obsolete (RBC-Ukraine)Unchecked AI: From self-replicating models to real-world harm (The Business Standard)Chine : comment des agents IA « voyous » ont échappé au contrôle de leurs créateurs (L'Opinion)