The AI Safety Bargain: Holding Back Models, Rushing Out Agents

The Debate

A company can decide a model is not ready for release—and, almost at the same moment, put AI agents to work around the clock. That is the awkward contrast at the center of OpenAI’s recent announcements: it scrapped the planned launch of GPT-6.1 Astra over safety concerns, then unveiled Dots, agents designed to pursue user goals across apps. Is that a sensible distinction between a model that failed a safety bar and a product built with limits, or is the industry applying caution...