AI application security focuses on the system around the model, not only on the model itself.

Important themes:

- prompts can contain hidden instructions or internal policy text
- retrieval systems can mix trusted and untrusted context
- outputs can leak more than the developer intended
- guardrails are often prompt-level suggestions rather than hard enforcement
- evaluation and grading flows create their own trust boundaries

For secure design, developers should separate public study content from internal-only content whenever possible.
