MasDrift: Benchmarking Authorization Preservation Across Multi-Agent Architectures
arXiv:2608.07556v3 Announce Type: replace-cross Abstract: Multi-agent systems (MAS) decompose long-horizon tasks across supervisors and subagents, but delegated goals do not necessarily carry their original authorization boundaries. Existing safety benchmarks mainly study adversarial compromise, while work on constraint drift lacks controlled architecture-level evaluation. We introduce MasDrift, a benchmark of 600 benign productivity tasks across eight domains. Each task pairs required work wit…