Introduction to AI Security Risks

AI Security is quickly moving from a niche research topic to a practical concern, as companies building and deploying AI systems are now facing real-world attacks and failure modes. This talk explores AI Security through a practical, attacker-focused lens. We’ll look at vulnerabilities that emerge at different stages, such as data poisoning, model extraction, and newer threats unique to large language models, such as prompt injection and tool misuse. We’ll also trace how the field has evolved, from early adversarial machine learning research to today’s complex ecosystem of generative models interacting with tools, APIs, and users in open-ended environments.

Finally, we’ll connect theory to practice by examining what AI security work looks like in industry today: the kinds of problems teams actually face, the roles being created, and the skills that are becoming essential as AI systems grow more autonomous and interconnected.

The goal is to give you a clear mental model of how modern AI systems fail, and how to think like the people trying to break (and defend) them.

Note: All times are shown in Nepal Time (NPT).