Adversarial Machine Learning: Attacking AMSI Providers

In this deep technical dive from the DEF CON Safe Mode AI Village, researcher Rob examines adversarial evasion techniques targeting machine-learning-based Antimalware Scan Interface (AMSI) providers. Opening at 00:01, Rob explores the fundamental mathematics behind AI-driven endpoint detection engines and how they classify suspicious execution payloads.

At 22:02, the talk contrasts rigid signature-based scanning with modern ML classifiers, demonstrating how slight structural modifications to file headers and magic bytes can break parser interpretation. At 44:02, Rob explores token-level adversarial perturbations, identifying specific byte sequences and keywords that Windows Defender tags as high-risk indicators and showing how to mask malicious scripts while preserving runtime functionality.

Moving past 01:06:00, the speaker demonstrates gradient-based evasion attacks against production ML models. Concluding at 01:28:01, Rob offers practical advice for malware defense engineers to harden ML classifiers against evasion by integrating adversarial examples into training pipelines. (AI summary — Gemini Flash)