Safety research reveals models shift internal states before risky actions. That's a big deal.

From the editor

Two threads run through today's briefing. First, safety: new research shows aligned training data can still produce misaligned models, while interpretability work finds measurable internal shifts preceding risky behavior. These aren't academic curiosities; they're the early warning systems the field has been hunting for. Second, efficiency: 97% of Claude Code tokens are cache re-reads, and Opus is overkill for most workloads. The pattern is clear. As capabilities sharpen, so does the pressure to run them cheaply and safely. Both frontiers are moving fast.