latent-shift

An independent researcher presents evidence that coherent context can shift LLMs into a different internal regime before producing output, bypassing surface-level safety filters. This suggests current alignment methods like RLHF may not be robust defenses.