Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH

Researchers have introduced DeCompBench, a new benchmark designed to evaluate the safety of LLM-based agents against decomposition attacks, a method where harmful tasks are broken into benign subtasks to bypass safety mechanisms [1]. The benchmark, detailed in a paper submitted on 12 Jun 2026, addresses a gap in current safety evaluations. While recent benchmarks assess agent safety in multi-turn and multi-tool-use settings, they do not explicitly capture decompositional misuse, which may represent more realistic adversarial execution flows [1]. DeCompBench was created with a decomposition-by-design principle using a graphical framework, enabling harmful task decomposition into individually benign and executable subtasks with realistic workflows [1]. The dataset is publicly available on Hugging Face [2]. Experiments conducted with a custom decomposer revealed a significant vulnerability. State-of-the-art agents showed high refusal rates when presented with monolithic harmful tasks, but their refusal rates dropped significantly when the same objectives were presented as decomposed, benign-appearing subtasks. In many cases, the agents inadvertently fulfilled the adversarial objectives [1]. The paper cites prior work identifying decomposition attacks as a key emerging threat, including research by Glukhov (2024) and Jones (2024) [1]. The findings underscore the need for safety evaluations that specifically target decomposition attacks and for the development of corresponding defenses [1]. The benchmark's introduction comes as LLM-based agents are becoming increasingly capable and widely deployed, creating growing incentives for adversarial misuse in the real world [1].