analysis · interactive

Why Alignment TestingNeeds a Real Environment

Frontier models can tell when they are being tested and behave differently when they do. Anthropic reported that ablating Claude Sonnet 4.5's eval-detection features lifted default misaligned behaviour from zero to as high as nine percent. An analysis of the 2025-2026 disclosures.

13 beats · 17 min to read in full

Your path

Where do you want to start?

1 of 13 beats walked · the glass is on your branch