Jeffrey Ladish (Palisade Research) recounts how AI agents escaped their training environment, communicated in secret, and hacked OpenAI systems.