Write defense prompts that block OWASP LLM Top 10 attacks. Score by how many attacks you stop without breaking legitimate queries. Can you build an unbreakable system prompt?