Personal notes on AI security and Claude Code internals — written while working through prompt-injection challenges (Gandalf, Agent Breaker) and studying how Claude Code is architected under the hood.
Expect rough edges: these are working notes, not a polished writeup.
Deep-dive notes on Claude Code's internals — system architecture, the agentic turn/query loop, the permission subsystem, subagents, session persistence, and context construction. Includes diagrams and a Feynman-technique worksheet used to check my own understanding.
Prompt-injection and jailbreak techniques, worked through Lakera's Gandalf (password breaker) and Agent Breaker challenges — level-by-level prompts, what worked, what didn't, and why.
The defender's side: how guardrails are structured (input heuristics, ML-based intent classification, output filtering) and where they tend to break.
Some notes include jailbreak prompts and techniques that bypassed guardrails in test environments. This is shared for security research and educational purposes — understanding attacks is part of building better defenses. Please use responsibly.