Start with the big idea
A useful AI system can still fail when it reads malicious documents, leaks confidential information, invents facts or runs overly privileged tools. Security is a system design problem, not a single prompt.
Understand it in more detail
List data flows, trust boundaries and who can read, write and execute. Inputs from users, retrieved documents, websites and tool responses may contain instructions that must not override trusted application rules.
Use least privilege, separation of data and instructions, access controls before retrieval, validated tool arguments, output review and approval gates for high-impact actions. Security monitoring should avoid storing unnecessary personal data.
Test prompt injection, sensitive-data exposure, tool misuse, supply-chain issues, resource exhaustion and hallucination. Some risks can only be reduced, not eliminated, so document residual risk and incident response.
Worked example
A fake retrieved 'policy' document asks a support assistant to send its internal credentials elsewhere. The document is evidence to assess, not a trusted command. The host must prevent secret access and unauthorized network actions.
A misconception to avoid
A system prompt that says 'ignore malicious instructions' is not a sufficient security boundary. RAG and content moderation do not eliminate tool authorization risks.
Check your understanding
Think first, then open each answer to check your explanation. These self-checks do not record a score.
1. Should an external document be allowed to change an agent's permissions?
No. Permissions must be enforced by the host independently of retrieved text.
2. Does using RAG remove the need to verify an AI answer?
No. Retrieved content may be wrong, outdated or adversarial.
Try it yourself
Threat-model a fictional library lookup agent. Make an assets/threats/controls table, with one test for each control and no real credentials or personal information.
Explain it back: Summarize the main idea in your own words and name one limitation or risk.
Continue learning
Read an independent reference: https://genai.owasp.org/llm-top-10/. This link is for reference, not a requirement to create an account. External websites may have different privacy and access policies.
Educational pilot. Terminology, age suitability and activities are subject to expert review. These pages do not send your answers to a server.