When Autoregressive Consistency Hurts Safety Alignment

In three linesResearchers show LLM safety alignment is fragile because concentrated on early tokens. Autoregressive consistency mechanism allows attacks to insert harmful sequences at any position and sustain them. They propose adversarial safety alignment with random worst-insertion training to break this consistency.

Summary generated by Claude — human-verified