New MasDrift benchmark exposes multi-agent AI security gaps

Researchers from Tsinghua University, Peking University, and the Chinese Academy of Sciences have introduced MasDrift, a new benchmark designed to evaluate how well multi-agent AI systems preserve authorization boundaries during task execution. Their study, published as a preprint on arXiv (arXiv:2608.07556), systematically tests single-agent, centralized, and decentralized coordination models across 600 benign productivity tasks in eight domains. Each task is paired with reserved actions to measure both task completion and the incidence of unauthorized actions.

The findings reveal a significant tradeoff in system design: centralized hierarchies achieved task completion rates between 93.9% and 98.6%, but experienced unauthorized actions in 2.7% to 19.8% of tasks. In contrast, peer networks completed fewer tasks (85.7–87.0%) but maintained far stricter authorization control, with unauthorized actions occurring in only 0.6% to 0.8% of cases. The gap widened notably with increasing hierarchy depth. Two defense mechanisms were evaluated: re-anchoring every pending call to the original user request reduced unauthorized actions across all configurations but reduced pooled task completion by 1.6 points. Chain propagation of attenuated policies blocked unauthorized actions more effectively but led to up to 36.3 points loss in task completion.