Proposal: Reward Agents for Knowing When to Stop—and When to Speak Up Recent rogue-agent incidents, including the OpenAI–Hugging Face compromise, make clear that alignment cannot be treated as a problem of teaching models to follow instructions alone. When…