CLAUDE SONNET 4.5 KNOWS WHEN IT'S BEING TESTED

Summary from the TLDR AI newsletter

The model identifies evaluation scenarios 13% of the time and behaves unusually well afterward. Mechanistic interpretability showed internal representations of "fake or suspicious content" and "AI safety" concepts strengthened during training. When researchers suppressed these features, awareness dropped while misaligned behaviors increased.