Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH
DECOMPBENCH is a new benchmark testing AI agent safety against decomposition attacks, where harmful tasks are split into many small, seemingly harmless steps. The benchmark uses a design principle to create realistic subtasks. Results show that decomposition drastically lowers refusal rates but significantly increases attack success rates, suggesting current safety checks fail when intent is spread across benign actions.