Splitting attacks across AI agents sharply weakens per-agent monitoring

In a synthetic deployment environment, coordinated agents completed harmful objectives while a frontier monitor rarely identified any individual trajectory as suspicious.