Abstract
AI workloads on Kubernetes inherit every cloud native vulnerability. And from GPU mega-clusters training hundred-million-dollar models, to MCP servers acting as universal adapters for autonomous agents, non-deterministic systems create a threat landscape that traditional cloud native security wasn't designed for.
In this talk, we examine the real-world attack surface of AI on Kubernetes: from NVIDIA container escapes and GPU memory side-channels, through tool poisoning and agent trust boundary failures, finishing with the governance gap that leaves critical services exposed.
Using the FINOS AI Readiness Governance Framework, we map controls to infrastructure, demonstrate potential gaps and highlight risk, and demonstrate that securing AI is still fundamentally Kubernetes security, with new crown jewels to protect.
Speaker
Andrew Martin
Founder and CEO @ControlPlane, CISO OpenUK, CNCF Security Advisory Group
Andrew is founder CEO at controlplane.io — securing regulated AI and cloud native systems. Hacking and hardening containers in production since 2014, he is co-author of Hacking Kubernetes (O’Reilly Media, 2022), an international conference and meetup speaker, and delivers training and whitepapers (SANS, Linux Foundation, O’Reilly, eBPF Foundation, Docker, Hashicorp). He is pro bono CISO for non-profit OpenUK.
His clients include Citibank, JPMC, Morgan Stanley, Google, Visa, S&P, PWC, BP, and various governmental departments.
Date
Tuesday Mar 17 / 10:35AM GMT ( 50 minutes )
Location
Mountbatten (6th Fl.)
