In 2025, U.S. retailers processed nearly $900 billion in returns. Of which, the National Retail Federation estimates 9% were fraudulent, putting return and concession fraud losses at $81 billion annually. Yet the security industry has largely handed this off to operations, while the operations side has lacked the security frameworks to treat it as an attack surface.

Now that gap is closing in the attacker's favor. Agentic AI has fundamentally changed the fraud economics: where return fraud once required a human caller, today's rings deploy voice bots generating over 1,000 AI-driven refund calls per day per chain. Visa's PERC research documented a 450% spike in dark-web mentions of AI agent fraud tooling in a single six-month window. OWASP's newly published Top 10 for Agentic Applications names Rogue Agents and Human-Agent Trust Exploitation as first-class threats - and they are showing up in refund queues, not just enterprise pipelines, making returns and concessions fraud not just an underbelly of larger corporate fraud, but also exposes the systems processing these as a major attack surface being exploited at an exponential scale by threat actors.

This session, co-presented by a senior security engineer and a senior program manager running customer returns and concessions at one of the world's largest e-commerce operations, walks through the full picture: fraud taxonomy, agentic threat classification, ML detection signals, event-driven response pipelines, and the physical and policy-layer countermeasures redefining triage at scale. This is not a workshop. Attendees leave with a mental model, a detection signal catalog, and a cross-functional operating picture ready to use.

Session Outline (50 minutes)

0-10 min: The $80B attack surface - types of fraud archetypes, how they scale, and the REKK case study from Telegram ad to international arrest

10-20 min: From human fraudster to AI agent -- the four classes of agentic commerce fraud and what changes when the counterparty on a refund claim is a bot

20-30 min: Detection signals - the feature catalog that distinguishes AI-driven fraud rings from legitimate customers, and how ML approaches are being applied to the returns domain

30-40 min: The event-driven fraud pipeline -- detection to automated response, Amazon's operational tooling in context, and how to avoid alert fatigue at scale

40-50 min: Countermeasures, the cross-functional operating model, the 2026-2027 watchlist, and Q&A