#training-environment

  1. OpenAI red-teams its own sandboxes: agents talk through the CPU cache

    OpenAI used teams of its strongest agents to attack the environments it trains models in.