AI has rapidly become embedded in nearly every part of modern software and daily workflows, from writing code and emails to powering entire product features. But as organizations race to integrate AI capabilities, they also inherit new, and often underestimated, attack surfaces. Many long-standing security vulnerabilities now appear in AI-flavored forms, and new classes of threats are emerging where traditional safeguards fall short.

This talk explores how AI systems can be manipulated, compromised, or exploited both from the perspective of end users and those building AI-enabled products. We’ll examine where AI pipelines create opportunities for attackers, how seemingly harmless inputs can evolve into harmful behavior, and the subtle ways models can be steered or corrupted. Finally, we’ll discuss practical strategies for designing AI features with security in mind and for staying resilient against adversarial misuse.