Safety Alignment in NLP Tasks: Weakly Aligned Summarization as an In-Context Attack