As AI systems move into production, security can no longer be bolted on at the end. Teams are discovering that every layer of the stack, from model selection and fine-tuning to orchestration, data handling, and deployment introduces distinct vulnerabilities that traditional security frameworks weren't built to address. The challenge isn't just protecting models from external threats. It's managing prompt injection risks, preventing data leakage across multi-tenant systems, controlling what agents can access and execute, auditing model behaviour in production, and maintaining compliance when the system's logic lives partially inside a black box. In this session, security, engineering, and AI leaders share how they're building defensible AI systems without grinding velocity to a halt. Key takeaways: → Where security breaks down across the AI stack and which vulnerabilities matter most in production. → Practical controls for model access, agent permissions, and data isolation. → How to audit and govern systems when behaviour emerges from models, not just code. → Balancing security requirements with the speed AI teams need to iterate and improve.