Blog and case studies
Notes from the frontier
First-principles essays, hands-on labs and case studies from the team building Freedom and Cosmipher AI Security.
Latest posts
How RAG Actually Works, and How Retrieval Gets PoisonedEmbeddings, chunking, vector indexes, hybrid search, and nine ways attackers turn retrieval against the model, with maths, real incidents, and a lab testing nine attacks against ten defences.Read on Substack
How AI Agents Use Tools, and Why Every Tool Call Is a Security BoundaryFrom first principles: the agent loop, function calling, MCP, and the seven ways a tool call goes wrong. With the risk math, real incidents, and a tested lab of seven attacks against ten defences.Read on Substack
How Transformers Actually Work: From First Principles to Modern LLMsFrom first principles to modern LLMs: attention worked by hand, why we divide by sqrt(d_k), positions, the KV cache, the O(n^2) cost, and a NumPy transformer you can train on a laptop.Read on Substack
Prompt Injection From First Principles: Why LLM Agents Can't Tell Instructions From DataWhy LLM agents mix up instructions and data, derived from first principles, with a runnable lab and a defence map.Read on Substack