We propose BAM-ICL, a novel budgeted adversarial manipulation hijacking attack framework for in-context learning.