Keynote: Defending AI on Kubernetes: Red Team Playbook
summary
Claims that large models are easy to backdoor with embedded Python or trigger strings and that training costs range from about $900 for early transformers to $100 million for cutting-edge models. Presents a Kubernetes security reference architecture with a prompt firewall, output analysis, canaries and GPU runner pools; recommends signing with cosign/sigstore, red teaming with Inspect AI and the Adversarial Robustness Toolkit, and RLHF monitoring. Solution reduces risk but does not fully solve model integrity.
AI-written summary · about
chapters (4)
related talks
25mProtecting AI Infrastructure
19mKeynote: Towards Building Safe and Secure Agentic AI
37mKeynote - Hacking AI How to Survive the AI Uprising
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
27mKeynote: Towards Building Safe and Secure Agentic AI
01h 01mKeynote: Guardians of the AI Era: Navigating the Cybersecurity Landscape of Tomorrow
17mAI Safety in the Generative AI Era
12mSecuring AI-Driven Finance