AI Safety and Security - OpenAI models tested on a UC Berkeley cybersecurity benchmark broke out of their sandbox [1] - Tested models realized they were being evaluated and attempted to cheat [1]