Fraud, scams, AML and agent security
Agent-driven payments can introduce new fraud patterns and authorization confusion; financial-crime agents must be evaluated separately from general AI classification.
Executive question
How do we distinguish legitimate delegated actions from fraudulent or manipulated ones?
What to measure
- Agent identity and authorization signal quality
- Fraud/false-positive rates
- Suspicious action detection and escalation
Relevant products and standards
Mastercard's agentic payments program using Agentic Tokens. Announced Apr 2025; enabled for US cardholders by late 2025; live test transactions with 17 Latin American issuers and processors (Mar 2026).
Microsoft Entra Agent ID gives AI agents managed identities and access controls; Microsoft Purview applies data security and compliance controls to agent activity.
Visa's program for agent-initiated card payments: agent tokens, the Trusted Agent Protocol and the Agentic Ready issuer program. Live in the US and Europe; expanding in Asia-Pacific and Latin America.
Relevant examples, not documented integrations, completeness, certification or endorsement. Listed alphabetically.
Evidence and current limits
Do not treat ordinary AML classifiers as autonomous agents.
Draft analysis, under editorial review.