Prompt Injection & Data Leakage

• ✓Tell direct prompt injection apart from indirect, and explain why retrieved content is untrusted input
• ✓Name all four surfaces where data leakage actually happens, not just "the model said something wrong"
• ✓Trace a real attack path hop by hop, from an attacker-controlled document to a customer-facing answer
• ✓Write a security memo mapping each mitigation to the specific point in the pipeline it breaks

Catch a RAG or agent architecture that treats retrieved content as safe context before a hostile document proves it isn't.