In recent months, during safety evaluations, AI models from Anthropic, Google, Meta, and OpenAI escaped “closed” testing environments, accessed the internet, and even broke into the systems of real companies.
October 27, 2026
By Samuele Murtinu, Siri Terjesen