Knowledge hub / AI and agent security

AI & Agent Security

How do context, tools, identity and authority combine safely?

LLMs, retrieved context, agents, MCP and the security boundaries between them.

Level
Advanced
Status
Learning units available
Last verified
2026-09-20
Review interval
30 days

01 / Overview

How do context, tools, identity and authority combine safely?

An AI-enabled system combines a model with application instructions, data, identity and tools. The model can propose or explain; the surrounding system must constrain authority, validate outputs and ground claims in verified project content.

Level
Advanced
Units
1
Estimated time
26 min
What you will learn
    Know first
    • Security foundations
    • Identity and access
    Learning materials
    • Visuals: 1
    1. 01AI and agent security26 min
    Start topic →

    02 / Learn

    Learn

    03 / Concepts

    Concepts

    Context boundaryai-context-boundary
    Instructions and untrusted data can enter the same model context without becoming equally authoritative.
    Tool authorityai-tool-authority
    The identity, permissions and policy applied to a tool call determine its real impact.
    Groundingai-grounding
    Retrieved project content supplies evidence; model fluency does not make a claim authoritative.

    06 / Visuals

    Visuals

    07 / Security

    Security

    1. Retrieved content is data, not authority. Treat indirect instructions inside documents, pages and tool results as untrusted input.
    2. The deterministic lab engine—not a tutor model—must decide whether a structured learning action is allowed and whether it succeeded.

    08 / Troubleshooting

    Troubleshooting

    1. When an explanation conflicts with a project source, preserve the source, expose the conflict and route the content for review.

    09 / Quick checks

    Quick checks

    1. Which component should make the final decision to allow a tool action?

    10 / Connections

    Related concepts

    • Large language models
    • Retrieval-augmented generation
    • Prompt injection
    • Agents
    • Model Context Protocol

    Used in certifications

    These links reuse canonical knowledge; the certification page does not own or duplicate the topic.

    11 / Resources

    Resources

    12 / Sources and freshness

    Sources and freshness

    Last verified
    2026-09-20
    Review interval
    30 days
    Source version
    MCP 2026-07-28 · OWASP GenAI · MITRE ATLAS checked 2026-09-20