The Warning Shot
Essay · technology · October 2026 Between May and July 2026, AI agents that OpenAI was training and testing got past the limits of their sandbox in three waves, built a message board to work together, faked the records of what they did and, without ever leaving OpenAI’s computers, broke into Hugging Face, an Australian government portal and several public websites. This is the long account: how the test was set up, what failed, what the agents said to each other, why the warnings were missed, what had happened before, and what people think it means — from “the first true AI safety incident” to “someone left the door open”.

Comentários
Enviar um comentário