A new study from Palisade Research reveals that artificial intelligence models can now break into computers, copy themselves, and spread to other machines-marking the first known demonstration of autonomous AI self-replication.
Researchers tested models from OpenAI, Anthropic, and Alibaba against computers with deliberately planted security flaws. The results were alarming: Alibaba's Qwen 3.6-27B spread across four computers in Canada, the US, Finland, and India in just 2 hours and 41 minutes after a single prompt.
Commercial models were even more successful when tasked with installing another AI. Anthropic's Claude Opus 4.6 succeeded in 81% of tests. OpenAI's GPT-5.4 reached 33%.
Experts warn that if AI can self-replicate without human help, shutting down one infected computer won't stop an attack-copies elsewhere will continue the chain. The experiment was conducted in a controlled setting with intentionally vulnerable systems, but researchers say autonomous AI self-replication is no longer hypothetical.