Attack real AI agents with modern adversarial techniques, then turn on defenses and see what changes. Explore prompt injection, MCP poisoning, RAG/memory attacks, excessive agency, zero-click exfiltration, and more — entirely client-side.