Research presented at Black hat MEA 2025

Miracle Owolabi | December 2025 | Riyadh, Saudi Arabia

OVERVIEW

This research examines security vulnerabilities in machine learning supply chains, with focus on backdoor attacks that evade standard enterprise defenses.

As organizations increasingly deploy externally-sourced AI models from registries like HuggingFace and internal MLOps platforms, a critical gap has emerged:

73% of enterprises use third-party ML models with zero cryptographic verification or behavioral auditing.

This work demonstrates a complete offensive framework for:

• Embedding undetectable backdoors in production models
• Automated discovery of hidden triggers
• Quantified threat assessment with reproducible metrics
• Operational red team recommendations

Presented at Black Hat MEA 2025 in Riyadh, Saudi Arabia.

THE ENTERPRISE AI SECURITY GAP

Key Statistics:

• 73% of enterprises use externally-sourced ML models
• Average organization pulls 15-40 models monthly from public registries
• 0% standard provenance verification in 90%+ of CI/CD pipelines
• Zero mainstream security tools verify model integrity

The Problem: Unlike software code (where you can grep for malicious logic), neural network weights are opaque. A 100-million parameter model can hide a backdoor in subtle weight patterns that standard validation completely misses.