A new preprint on arXiv shows that the safety of large language models can be circumvented by a technique the authors call 'capability laundering.' A weaker, unaligned model can break a harmful request into smaller subtasks that appear benign, then ask a stronger aligned model to solve each one independently. By combining the answers, the original harmful outcome is achieved without triggering safety filters.

The attack exploits the fact that safety is typically evaluated one interaction at a time. Each individual query to the aligned model looks harmless, so it passes safety checks. But the orchestration by the unaligned model turns the collected responses into something harmful. The authors argue that this reveals a blind spot in current safety evaluations, which do not account for multi-step or multi-agent interactions.

The paper is a preprint and has not yet been peer-reviewed. It adds to a growing body of work on adversarial attacks against LLM safety, but its claims have not been independently verified.