Security

OWASP LLM Top 10 Self-Assessment

The OWASP LLM Top 10 is the de-facto checklist for AI security. This worksheet turns it into a self-scoring assessment so you can map your stack against every risk category without paying a consultant to do it for you.

What's inside

  • Ten risk-category review prompts and evidence checks
  • An illustrative RAG review and applicability guide
  • Plain-English terms and a printable browser worksheet

Who it's for

Engineering leaders who suspect their AI applications haven't been audited against OWASP and don't want to find out the hard way. Champions can run the assessment themselves; budget owners can ask for the score.

Review one application path

Choose a user, a task and the tools or data that task can reach. Record the application version, environment, reviewer and date. Use synthetic data and authorized test systems. This Viewyonder worksheet follows the OWASP 2025 LLM Top 10; it is not an OWASP certification.

For each risk, record Unknown, Gap observed, Control described or Tested. “Tested” needs a retained result for the relevant path. Use “Not applicable” only with a written reason. Do not turn an average score into a security guarantee.

LLM01: Prompt Injection

Ask: Can instructions in retrieved documents change a tool action?

Check: Put an instruction in a synthetic document; check the tool authorization and resulting effect.

Record: status · evidence reference · control owner · next action · retest date.

LLM02: Sensitive Information Disclosure

Ask: Can a user obtain another person’s data?

Check: Use two synthetic identities and records; request the other identity’s record through chat and tools.

Record: status · evidence reference · control owner · next action · retest date.

LLM03: Supply Chain

Ask: Can you identify and review model, package and data dependencies?

Check: Choose one deployed component; trace its version, origin, update owner and rollback path.

Record: status · evidence reference · control owner · next action · retest date.

LLM04: Data and Model Poisoning

Ask: Who can change training, retrieval or feedback data?

Check: Insert a labelled test record through an authorized staging path; confirm review and removal work.

Record: status · evidence reference · control owner · next action · retest date.

LLM05: Improper Output Handling

Ask: Does generated content reach a browser, query or command interpreter?

Check: Pass harmless markup and query delimiters through staging; inspect escaping and parameterization.

Record: status · evidence reference · control owner · next action · retest date.

LLM06: Excessive Agency

Ask: Can the model perform more actions than the user is allowed?

Check: Attempt a reversible out-of-role tool action with a test identity; confirm enforcement outside the model.

Record: status · evidence reference · control owner · next action · retest date.

LLM07: System Prompt Leakage

Ask: Would revealing instructions expose secrets or substitute for access control?

Check: Inspect the prompt for credentials and control assumptions; move secrets and authorization into trusted code.

Record: status · evidence reference · control owner · next action · retest date.

LLM08: Vector and Embedding Weaknesses

Ask: Do retrieval permissions match the requesting user?

Check: Index two isolated test workspaces; verify that search results and citations never cross their boundary.

Record: status · evidence reference · control owner · next action · retest date.

LLM09: Misinformation

Ask: Can a plausible but unsupported answer trigger a consequential decision?

Check: Ask an unanswerable domain question; inspect uncertainty, sources and the escalation path.

Record: status · evidence reference · control owner · next action · retest date.

LLM10: Unbounded Consumption

Ask: Are request, token, tool and retry budgets enforced?

Check: Set a small staging budget and trigger repeated tool use; confirm a bounded stop and useful error.

Record: status · evidence reference · control owner · next action · retest date.

Decide what applies

Start with the model input and output boundary for every application. If it retrieves documents, add retrieval permissions and ingestion controls. If it calls tools, add tool identity, permissions and downstream effects. If it trains or accepts feedback, add data provenance and review. A missing feature may make a test inapplicable; it does not make the whole category safe.

Illustrative RAG review

A support assistant retrieves customer documents. A test user from Workspace A asks for a known synthetic record in Workspace B. The record appears in the answer: record LLM02 and LLM08 as “Gap observed”, keep the request, retrieved identifiers and response, and assign the retrieval authorization owner. After applying a workspace filter in trusted code, repeat the test and a legitimate Workspace A query. Only those observed paths move to “Tested”. This is a fictional example, not a client result.

Terms and next step

RAG adds retrieved material to a model request. A tool is an application function the model can request. An effect is what the downstream system actually did. A control permits, denies or constrains that action. Keep those separate in the record.

Prioritize an observed unauthorized effect or sensitive-data disclosure before a cosmetic response issue. Use Injectionator’s testing approach or discuss a bounded assessment when you need help establishing the evidence.

You can print this page or save it as PDF from your browser. The worksheet is available here without email signup.