AI on Kubernetes is a hidden attack surface

AI Workloads on Kubernetes inherit every Cloud Native vulnerability and then introduce a whole extra class of threats that traditional tooling wasnt designed to catch, and isn't quick enough to defend against.

From GPU mega-clusters training million dollar models, to MCP servers acting as a universal adapter for autonomous agents, the infrastructure running today's AI is riddled with attack surfaces most teams haven't even considered yet. The vulnerabilities are real, they are exploitable in production and the blast radius when they go wrong are huge.

This session draws on hands-on experience securing AI infrastructure at scale to walk through the real attack surface. NVIDIA container escapes, GPU memory side-channels that leak models weights and training data, tool poisoning attacks on agentic systems and agent trust boundary failures that turn your autonomous pipeline into an uncontrolled execution environment.

Using the FINOS (Fintech Open Source Foundation) AI Readiness Framework as a starting point we'll discuss where controls exist, where the gaps are, and how to communicate that risk to leadership and boards. At the end of the day, securing AI is still fundamental cloud native security, just with more enticing treasure and greater consequences. This isn't a guess about the future of AI, it's a practical view of where the active threats already exist today.