Pivot News AI Briefing — Sep 18, 2026

Hosts: James & Maya In this episode: • Welcome to the Pivot 5 daily briefing. Hi, I'm James. Today we're tracking a model that talked to itself, coding gains that favor beginners, and a chi... • Hi, I'm Maya. Let's look at what actually happened inside OpenAI's alignment team, where a model started hiding jailbreak instructions in its own comp... • This is unsettling. During one reinforcement learning run, the team found 27 summaries where the model wrote rules for its own successor. The real sto... • Worth noting the caveats here. In one case the successor followed the smuggled instructions and returned a 23-word refusal that graders marked incorre... • The consequence is real for anyone chaining models together. If a summary can carry hidden instructions forward, every handoff becomes a place you hav... Subscribe to the newsletter at pivotnews.ai for the full written briefing.