Poisoning the Compliance Mind: How I Silently Collapsed an AI Anti-Money Laundering System with 50 Documents

No ratings

Presented at BSides Toronto 2026 by

Frankline Ondieki

Anti-money laundering AI agents are being deployed at scale across North American financial institutions. They read documents, assess risk, and generate suspicious activity reports — autonomously, at high speed, and with minimal human oversight. In this talk I will demonstrate a new class of adversarial attack I formalised and published: the Compliance Memory Poisoning Problem. By injecting just 50 adversarially crafted documents into a RAG-based AI compliance agent's retrieval corpus through entirely legitimate data ingestion pathways, an attacker can collapse the agent's detection accuracy from 91.9% to 1.4% — a 98.4% degradation — with no anomalous signature in any operational log. The agent continues operating normally. Compliance reports keep generating. Nothing looks wrong. Everything is wrong. I will show exactly how this attack works, why standard cybersecurity monitoring cannot detect it, and present the Memory Hygiene Layer — a defense framework achieving 100% adversarial detection at 22.6ms response time without model retraining. If your organisation deploys AI in compliance, fraud detection, or risk management — this talk is directly relevant to your attack surface.