"Alignment Forecasting: Predicting Misalignment From Training Data" This paper has an AI inspect the training data and forecast whether fine-tuning will increase behaviors like deception, sycophancy, or sabotage. Their decomposed forecaster reaches 0.801 at predicting whether fine-tuning will trigger misalignment, versus 0.48~0.65 for directly asking frontier LLMs to make the same prediction, and its signals can also help filter risky training examples. https://alphaxiv.org/abs/2609.35805