docs/problem_framing.md

The Problem: Static Defenses vs Adaptive Generative Fraud

In the traditional payment security paradigm, defenses are static. Machine learning models and rule-based thresholds are trained on historical fraud data, effectively fighting the last war.

With the proliferation of Generative AI, malicious actors no longer need to manually craft individual phishing messages or discover static evasion thresholds by trial and error. Instead, they can deploy autonomous, adaptive agents that iterate on attacks in real-time.

The GenAI Advantage

1. Mass Personalization (Attacks 1 & 2): LLMs can synthesize deeply personal context scraped from breaches, turning generic spray-and-pray phishing into highly targeted spear-phishing at scale.
2. Adversarial Camouflage (Attack 3): An LLM-driven adversary can test rule boundaries—such as transaction velocity limits—and iteratively restructure its fraudulent transfers (e.g. smurfing/structuring) to evade detection dynamically.
3. Agentic Vulnerabilities (Attack 4): As consumers and merchants adopt autonomous AI agents to manage checkout and purchasing pipelines, a new attack vector emerges: Prompt Injection. Attackers can poison public product descriptions or reviews to hijack the autonomous purchasing pipeline.

The Solution: A Closed-Loop AI Defense Lab

To defend against adaptive generative fraud, our defenses must be equally generative. The AI Defense Lab creates a synthetic testing ground where we:

• Identify novel, emerging GenAI vectors.
• Generate massive synthetic datasets of these attacks before they hit the wild.
• Defend by retraining our production models on the synthetic data, closing the loop.

This project proves that you can pre-emptively discover and patch vulnerabilities in your ML detection systems by proactively generating the attacks you fear.