Inside the LLM | Understanding AI & the Mechanics of Modern Attacks

**Executive Summary:** This report analyzes how LLM inference pipeline stages—tokenization, embeddings, positional encoding, and self-attention—create practical attack surfaces (e.g., filter bypass via tokenization boundaries, gradient-based embedding shifts, context-window chunking, and attention-hijacking adversarial suffixes) and surveys mitigations such as randomized smoothing, suffix filtering, and adversarial training.