Federico Maggi - [Keynote] Secure AI Starts with Solid Compute Foundations

AI Stack Under the Hood

Trend Micro/Federico Maggi’s No Hat keynote traces modern AI from HTTP-authenticated microservices down through containers, VMs, and servers to LLM frameworks that iteratively predict next tokens trained on text, code, and security-relevant content. Agentic systems add orchestration across LLMs, APIs, and MCP-style tools.

Compute Foundations for LLMs

Transformers embed tokens, apply layered numerical transformations, and output probability distributions over vocabulary. Training trillions of parameters demands specialized scale beyond general-purpose CPUs—GPUs parallelize locally; ASICs like AWS Trainium 2 use systolic arrays fed directly from high-bandwidth memory without traditional instruction streams.

Accelerator Security Boundary

Trainium cards include a management SOC controlling reset, PCIe setup, and security configuration. Maggi’s threat model assumes cloud tenants may run as root with custom drivers, so the management microcontroller must enforce isolation at the machine level—secure AI, he argues, starts with trustworthy compute hardware beneath models and agents.