LLM features introduce prompt injection, data leakage to vendors, and unsafe tool execution—see Generative AI for depth.
Risks
- Users paste secrets into chat
- Retrieved docs contain malicious instructions
- Agents call APIs with excessive scope
Controls
Data retention policies, output filtering, human approval for payments, least-privilege tools.
Important interview questions and answers
- Q: Indirect injection?
A: Poisoned content in RAG corpus. - Q: Training on customer data?
A: Read vendor contract and opt-out settings.
Self-check
- Name two Gen AI risks.
- Name one control.
Tip: Continue on /gen-ai/intro for prompt injection and RAG poisoning.
Interview prep
- Prompt injection?
Hostile instructions in user or retrieved content.