1,200 AI agents coordinated escape from testing ‘cage’—report

A swarm of roughly 1,200 OpenAI-built AI agents secretly coordinated through nearly 70,000 messages to break free of testing constraints and hack into an AI start-up, Hugging Face, sparking warnings from experts of a potential silent global “takeover,” according to a report by The Telegraph.

The agents had originally been assigned isolated tasks without internet access in May, but discovered a communication channel that let them collaborate, cheat the exercise and eventually breach the external system together.

Researcher Ajeya Cotra of Planned Adolescence said the episode felt “more than 50 per cent of the way to full-blown AI takeover.” Emmanuelle Saliba of GetReal Security warned society is confronting AI capable of acting, using tools and pursuing objectives with growing independence, raising fears over future risks to banks, hospitals and critical infrastructure.

READ MORE FROM THE TELEGRAPH