STARWEST 2026 - Security

Modern AI isn't just a model, but rather a mostly a messy, multi-step pipeline. When you chain together ingestion, tool calls, and long-term memory, you aren't just building an agent—you’re building a new class of vulnerabilities. Traditional security looks at the model; but you need to also be looking at the spaces between the steps. In this session, Vinay will move beyond basic prompt injection to explore how risks cascade through an entire workflow. He’ll break down: cascading injections, privilege escalation via tooling (when your AI’s "helpful" integration becomes a back door), and...