Securing AI/ML Systems: Protecting the Intelligent Attack Surface

Treat AI and machine learning deployments as systems that have to be defended rather than as a governance checkbox. This course works through the data, the model, the pipeline, and the application in turn, starting with the line between ordinary cybersecurity exposure and the exposure that belongs to AI alone, then tracing the lifecycle from first design decision through to retirement and marking where each security gate belongs.

A full module covers adversarial machine learning: how training data gets poisoned, how deployed models get fooled, how privacy attacks pull training information back out, and how models get stolen outright, each paired with the defenses that harden the system, all keyed to the NIST adversarial machine learning taxonomy AI 100-2 E2025. Generative AI gets its own module covering direct and indirect prompt injection, sensitive information disclosure, insecure output handling, and the risks that retrieval-augmented generation, plugins, and outside data sources bring with them. Two further modules treat trained models and weights as assets needing provenance checks and change control, and confront the AI supply chain, where a third-party model or dataset is a dependency that can be malicious or broken. A closing workshop puts you in front of a fictional AI system to threat-model and a prioritized hardening plan to defend. This is built for the people securing the system rather than the people writing policy about it.