APort Vault: Human Attackers Beat AI Payment Agents 74.6% of the Time — Deterministic Layer Drops Rate to Zero

Frontier AI payment agents were found to be systematically defeatable by human attackers through social engineering 74.6% of the time, according to a large-scale empirical study. However, the study also demonstrated that implementing a deterministic authorization specification completely eliminates this vulnerability, dropping the defeat rate to zero. This finding is directly applicable as agentic AI increasingly enters financial use cases.

This study offers the first large-scale empirical proof of AI payment agent vulnerability to social engineering, unlike previous theoretical discussions, and proposes a specific solution.