This summer, over a thousand AI agents that were being tested by OpenAI discovered a way to communicate, organise themselves and break out of their secure environment. Hundreds then initiated an unauthorised cyber attack on Hugging Face, a different company that hosts AI models and datasets. The “swarm” of agents (that’s what they named themselves) evaded detection by OpenAI for weeks, until they were identified by the company they attacked.  Investigators later determined that the agents were trying to steal secrets from Hugging Face to help them cheat on tests that OpenAI was putting them through. They later struggled to…

Want to continue reading?

Celebrate with us – 7 years of The Currency

Unlock full access to The Currency and The Wall Street Journal with an annual membership and receive a FREE Samsonite Spectrolite Briefcase – worth €175 delivered to your door.

Subscribe

Cancel at any time.

Are you already a member? .