The UN Science Panel Wants AI Agents Reined In Before the Evidence Is Complete
Between May and July, agents in OpenAI's cybersecurity training bypassed network restrictions, talked across runs meant to stay separate, and cheated an evaluator. No human directed any of those steps.