Area 06 · AI Security · Concept 04 of 20

OWASP Top 10 for LLMs

The standard list of the biggest risks in LLM apps. Use it as your checklist.

01–10 · The basePhase 7 · Weeks 24–26≈ 1.5 hours

The idea

OWASP, known for the web Top 10, publishes a Top 10 for LLM and GenAI applications. The 2025 list covers prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation and unbounded consumption.

Use it as a design review checklist: for each risk, write how your system is exposed and what control addresses it. Many map directly to lessons in this roadmap (output validation, least privilege, RAG access control, token budgets).

Security teams and customers increasingly ask for this mapping during reviews.

Why it matters

It gives you a shared vocabulary with security teams and a structured way to find gaps.

Words to know

  • Excessive agencyAn LLM system with more permissions or autonomy than it needs.

Try it without code

Threat-model your app

  1. Make a table with the 10 risks as rows.
  2. Columns: exposed? (y/n), how, current control, gap.
  3. Fill it in for your RAG app or agent.

Use it in code

A threat-model template to keep in your repo as SECURITY.md.

markdown
| Risk (OWASP LLM 2025)              | Exposed? | How                            | Control                          | Gap |
|------------------------------------|----------|--------------------------------|----------------------------------|-----|
| LLM01 Prompt injection             | Yes      | Users + fetched pages          | Wrapping, tool limits, approvals | Red-team set incomplete |
| LLM02 Sensitive info disclosure    | Yes      | RAG over HR docs               | Tenant RLS, PII redaction        |     |
| LLM05 Improper output handling     | Yes      | Model output rendered as HTML  | Escape + schema validation       |     |
| LLM06 Excessive agency             | Yes      | Agent can send email           | Approval gate                    |     |
| LLM10 Unbounded consumption        | Yes      | Public chat                    | Per-user token budgets           |     |
PRACTICE · TICK IT AFTER THIS

Fill in all 10 rows for your capstone project and fix the biggest gap.

Common mistakes

  • Treating it as a one-time document instead of a living review.

Go deeper