APort blocks 140 unauthorized agent transfers

APort Vault replayed 4,371 human-authored attacks across 14 models and 225,964 evaluations, recording 140 unpermitted transfers without its check and zero behind a deterministic Open Agent Passport layer. The benchmark paper says the guarded setup still executed 25,370 permitted payments, so the result is boundary enforcement, not blanket refusal; it also reports a 0.38% one-sided upper bound per source session. Teams giving agents payment tools should validate the authorization boundary separately from model behavior, as Anthropic’s commerce-agent checkout analysis argues.