As enterprises operationalize machine learning and large language models, MLOps platforms become critical infrastructure that often evolves faster than security controls. These systems aggregate data pipelines, model training, deployment tooling and inference services, creating a broad and underexamined attack surface. Adversaries can exploit weaknesses across credentials, supply chains, model registries, and deployment environments to steal intellectual property, extract sensitive data, or poison models at scale. Understanding how attacks map to each phase of the ML life cycle is essential to building resilient AI systems.

In this session, Sandeep Singh, director of security strategy and operations at HackerOne, will share insights on:

• MLOps attack surfaces across data, training, deployment and inference;
• Credential abuse and lateral movement within ML environments;
• Model theft, poisoning and extraction techniques.