• Two main leak paths 2:18:
    1. The AI system has access to a database with sensitive information and leaks it into responses.
    2. The model was fine-tuned on data containing private information, which can resurface in later generations.
  • This risk list is not complete: for a fuller catalog, the video points to the OWASP Top-10 LLM vulnerabilities list. 2:34

Mitigations covered later: constraining model data access, sanitizing fine-tuning data (product design choices), and input screening (guardrails).