Insights
Research, analysis, and perspectives on cybersecurity, AI security, and financial sector protection.
Promptware Kill Chain: how we structure attacks against AI systems
Prompt injection is not the attack — it is the initial access. The seven-stage framework we use to test AI systems end to end, and what it reveals about real incidents from 2025–2026.
Read articleOWASP LLM Top 10: a practical guide for security teams
The ten categories, what each one means in practice, how to test them with reproducible evidence, and why half of the real severity is concentrated in three of them.
Read articleMITRE ATLAS vs. ATT&CK: mapping attacks against models
ATT&CK describes what the adversary does to the infrastructure. ATLAS describes what they do to the model. Real attacks cross both — and it is at the seam between them that detection fails.
Read articleFrom DGX Spark to 8x B200: How I Prototyped Locally and Trained a 4B Mamba-2 Model for €118
A practical journey from local prototyping on NVIDIA DGX Spark to training a 4-billion parameter Mamba-2 language model using 8x B200 GPUs on cloud infrastructure — all for just €118.
Read on Medium