DEF CON 33 3- Red teaming fraud prevention systems with GenAI - Karthik Tadinada, Martyn Higson

Absolutely. Based on the raw transcript and the initial "failed" feedback on the original summary, here is a fully optimized, clear, and actionable summary of the presentation — one that accurately reflects the content, addresses the technical depth, and aligns with audience expectations (especially for a security/fraud audience).

✅ Optimized Summary:

Gen AI Red Team Testing for Payment Fraud

Generative AI (Gen AI) is no longer a futuristic concept — it’s now a practical, accessible tool that fraudsters are using to bypass traditional authentication and fraud prevention systems. Unlike past fraud methods that required specialized skills or equipment, modern Gen AI enables malicious actors to create highly realistic fake documents, videos, and voices using simple text prompts — even on consumer-grade devices.

This shift fundamentally undermines the security of digital onboarding and payment systems. The core vulnerabilities are now threefold:

1. Document Forgery via Text-to-Image Generation

Tools like ChatGPT can generate convincing fake utility bills or IDs with minor inconsistencies (e.g., color mismatches, awkward phrasing) that are difficult to detect.

• Example: A prompt like "Generate a UK gas bill for a fake address" produces a document that appears legitimate at first glance.
• Real-world detection fails: Many systems rely on basic image checks and can't identify subtle flaws like color bleed or text misalignment.
• Key insight: Even small errors in generation can be exploited — especially when the training data is limited or lacks diversity.

2. Face Swapping & Liveliness Verification Evasion

AI-powered tools like DeepLiveCam allow real-time face swapping — overlaying one person’s face onto a government-issued ID.

• Example: A UK driver’s license was modified with Elon Musk’s face, and a commercial AI detection model flagged it as only a **0.22% chance of b