STRIDE still works when a language model sits in the middle — if you treat the model as an untrusted component.
Diagram, trust boundaries, controls on the wire, and a human gate on the steps that must not be casual.
Baseline versus protected profile, 17/17 PASS, and the limits of that number. From the MSc thesis.
A replayable suite, not a chat screenshot: layer hit, HTTP status, latency, block reason.
Security as middleware: an OpenAI-compatible reverse proxy for LLM traffic.
If the control plane is a Base URL, notebooks, chatbots and batch jobs share the same pipeline.
How Fast Gradient Sign Method generates adversarial examples — a lab note from the MSc in AI.
Gradient-based attacks, visible perturbations, and a first measure of model robustness.
Short write-ups in English, roughly weekly: AI security, SSDLC, and the boring controls that actually ship. The thesis pieces on the AI Security Gateway stay here as the technical backbone.