Multi-agent systems (MAS) break long-horizon tasks into subtasks handled by supervisors and subagents. But when a goal is delegated, the authorization attached to the original task does not necessarily travel with it. That gap can lead to subagents acting with broader or narrower permissions than intended.
To address this, a new arXiv preprint introduces MasDrift, a benchmark for measuring authorization preservation across multi-agent architectures. The paper argues that existing safety benchmarks focus primarily on adversarial attacks, whereas MasDrift targets the more subtle problem of authorization boundaries being lost or altered during delegation.
The abstract does not provide experimental results or specific architecture comparisons, so the contribution at this stage is the benchmark itself. The authors position it as a complement to current safety evaluations, which tend to overlook non-adversarial authorization drift.