AI scammers win 46% compliance in blinded trial

An LLM romance-scam agent won 46% request compliance in a week-long blinded study, versus 18% for human operators. The USENIX Security 2026 paper also found greater participant trust in the agent and a 0.0% detection rate for romance-baiting dialogues across popular safety filters; the work included interviews with 145 scam insiders and five victims. For teams building the autonomous workflows described in Google’s agent-factory economics, the defensive lesson is stark: moderation must inspect long-horizon intent and relationship patterns, not merely wait for a visibly fraudulent final request.