← All news and guides
Security

AI Security Review Template: OWASP LLM Top 10 Mapped to Evidence

A reusable AI security review template: each OWASP LLM Top 10 risk gets a review question, evidence to attach and a status, plus additions for agents and MCP.

Mono Team ·

A flat file cabinet with one wide drawer pulled open, showing neat blank folders in a dim archive room lit by a single window.

Short answer: an AI security review template is a table with one row per OWASP LLM Top 10 risk, and four things you fill in for each AI system: the review question, the control you run, the evidence that proves it, and a dated status. The evidence column is what matters. A row that says "we handle prompt injection" is a claim. A row that links to a test set, its results and a date is something a customer or auditor can check. Run the review before a feature ships and again when the model, the provider, the data sources or the tools it can call change. This post gives you the template, the evidence to attach for each risk, and the additions for agents and MCP tools. It is mapped to the OWASP Top 10 for LLM Applications 2025. It is not "OWASP compliant", because OWASP publishes no certification.

What is an AI security review template, and when do you use it?

It is a repeatable record of how one AI system could be attacked or misused, what you do about each risk, and how you know the control works. You fill in one copy per system or feature, not one for the whole company. A support chatbot, a document summariser and a coding agent have different risks and need separate copies.

Run it at these moments:

  • Before first release of any feature that sends data to a model or acts on a model's output.
  • When anything underneath changes: a new model or provider, a new data source for retrieval, a new tool or permission for an agent, or a new customer type.
  • When a buyer sends a security review that has an AI section. The finished template is most of your answer. See our guide to answering the AI section of a questionnaire.
  • On a calendar, at least once a year, so stale evidence gets noticed.

How do you scope the review before you score anything?

A clean steel workbench seen from a low angle with a pair of calipers measuring a small machined metal part, soft side light, shallow depth of field, no text and no faces.

Most weak reviews fail at scope. People score risks for a model they have not described. Fill in this sheet first and attach a one-page data flow diagram.

Scope fieldWhat to write
System and ownerWhat to write: The feature name and one named person who can answer for it
Models and providersWhat to write: Each model, its version, who hosts it and where it runs
Data sourcesWhat to write: Everything that enters a prompt: user text, uploads, retrieved documents, tool results, web pages, email
Who can reach itWhat to write: Roles, tenants and whether the public can use it
What it can doWhat to write: Read only, write, send messages, run code, spend money, call other services
Where output goesWhat to write: A screen, a database, another system, a customer email, a code runner
Logging and retentionWhat to write: What is stored, for how long, and who can read it
Review detailsWhat to write: Reviewer, date, and the change that triggers the next review

The "what it can do" and "where output goes" rows set how serious everything else is. A chatbot that only displays text has a small blast radius. The same model wired to your CRM and email has a large one.

How do the OWASP LLM Top 10 risks map to review questions and evidence?

The OWASP Top 10 for LLM Applications 2025 lists ten risks, LLM01 to LLM10. The table below turns each into one review question and the evidence that answers it. The OWASP project publishes detail and mitigations for every item, so read those pages for the risks that matter most to you.

OWASP riskReview questionEvidence to attach
LLM01 Prompt InjectionReview question: Can text from a user, document, web page or tool result change what the model does, and what is the worst it could then do?Evidence to attach: Injection test set with dated results; list of untrusted content sources; the permissions the model holds
LLM02 Sensitive Information DisclosureReview question: What sensitive data can the model see, and can one user's data reach another user or a log?Evidence to attach: Data classification per input source; tenant isolation test; redaction rules; retention settings
LLM03 Supply ChainReview question: Which models, libraries, datasets, plug-ins and tool servers does it depend on, and who approves changes?Evidence to attach: Inventory with versions and provider terms; version pinning; update and approval process
LLM04 Data and Model PoisoningReview question: Can anyone alter training, fine-tuning or embedding data, or the documents we retrieve?Evidence to attach: Data lineage; list of who can write to each corpus; review step for new sources
LLM05 Improper Output HandlingReview question: Is model output checked before it reaches a browser, database, shell or API?Evidence to attach: Output schemas; escaping and parameterised queries in code; tests with hostile output
LLM06 Excessive AgencyReview question: Which tools can it call, with whose permissions, and which actions need a human?Evidence to attach: Tool list with permission per tool; approval rules; spend and rate limits; kill switch test
LLM07 System Prompt LeakageReview question: If the system prompt is read, does anything sensitive leak, and does security depend on it staying secret?Evidence to attach: Prompt review showing no keys or access rules in it; access control enforced in code
LLM08 Vector and Embedding WeaknessesReview question: Does retrieval respect who is asking, and who can write to the vector store?Evidence to attach: Per-user retrieval filter; test that user A cannot retrieve user B's content; write-access list
LLM09 MisinformationReview question: What happens when the model is confidently wrong, and who is harmed?Evidence to attach: Evaluation results on your own cases; grounding approach; human review rule for high-stakes outputs
LLM10 Unbounded ConsumptionReview question: What stops one user, loop or attacker from running up compute, cost or delay?Evidence to attach: Rate limits; token and time caps; per-tenant budgets; billing alerts

What does good evidence look like for each group of risks?

Input and instruction risks: LLM01 and LLM07

Keep a test set of injection attempts and run it whenever the prompt, model or tools change. Include direct attempts, where a user types an override, and indirect ones, where instructions are hidden in a document, web page, email or tool result. A pass means "no failure found on this set on this date". It does not prove safety, so record it that way.

The stronger control is limiting what a fooled model can do, which is LLM06. For LLM07, assume the system prompt can be extracted. Keep keys, connection strings and access rules out of it. Enforce access in code.

Data risks: LLM02, LLM04 and LLM08

Draw the path of each data source into the prompt and mark its sensitivity. Then test the edges. Create two test tenants, put a unique marker in one tenant's documents and confirm the other cannot retrieve it through the assistant. Check that logs and traces do not store prompts in a place more people can read than the original data. Check your provider's data-use and retention terms against what you tell customers.

For poisoning, the question is who can write. Anyone who can add a document to your retrieval corpus can influence answers. List those people and systems, and add a review step for new sources. Retrieval filters should use the caller's identity at query time, in the retrieval layer, not an instruction in the prompt.

Output and action risks: LLM05 and LLM06

Treat model output as untrusted input. Escape it before it is rendered in a browser. Use parameterised queries, never string-built SQL from model text. Validate structured output against a schema. Never pass it to a shell or an evaluator. Then list every tool the model can call and apply least privilege to each:

  • Run tools with the end user's permissions where you can, not a shared service account.
  • Require human approval for actions that are hard to undo: sending, paying, deleting, publishing.
  • Cap spend, call counts and loop depth per request.
  • Test the kill switch. Record how long it takes to disable the feature.

Dependency risks: LLM03

Your inventory should list the model and version, the provider and its terms, packages, datasets, plug-ins and any MCP servers. Pin versions. Give each addition an owner who approves updates. Developer tools count too: record which AI coding assistants engineers use and what each does with your code, and date every vendor claim you rely on.

Reliability and cost risks: LLM09 and LLM10

Build an evaluation set from real cases in your domain and run it on every model or prompt change. Decide which outputs need a human check before they reach a customer. For consumption, set per-user and per-tenant limits, hard token and time caps, and an alert on spend. A loop or a retry storm is as likely as an attacker.

What changes when the system is an agent or uses MCP tools?

A wall-mounted electrical distribution cabinet opened to show tidy bundled cables and rows of identical breakers in a dim utility room, soft overhead light, no labels or text legible.

Agents raise the stakes because they act. OWASP published a separate Top 10 for Agentic Applications on 9 December 2025. Its ten items are Agent Goal Hijack (ASI01), Tool Misuse and Exploitation (ASI02), Identity and Privilege Abuse (ASI03), Agentic Supply Chain Vulnerabilities (ASI04), Unexpected Code Execution (ASI05), Memory and Context Poisoning (ASI06), Insecure Inter-Agent Communication (ASI07), Cascading Failures (ASI08), Human-Agent Trust Exploitation (ASI09) and Rogue Agents (ASI10). For an agent, add a second block of rows to the template using these.

If the agent connects to tools through the Model Context Protocol (MCP), add rows for what the 2026-07-28 specification leaves to you. Authorization is optional in the spec, so record whether each server authenticates callers and checks that tokens were issued for it. The spec also recommends a human in the loop who can deny tool calls, so record the approval rule. Tool servers are software with their own published CVEs, which is why the inventory needs a version for each one. Our post on MCP server security risks covers the spec's token rules, the CVE records and least privilege for agents in detail.

Agent additions to the template: a tool inventory with owner and version, the scopes and credentials each tool holds, how tokens are stored, the approval rule per action, and an action log you can replay. Buyers ask about these too: see new questionnaire questions about AI agents.

How do you score the review and decide what blocks a release?

Use four statuses and apply the same rules to every row.

StatusRule
MetRule: Only if you can link an artefact with a date: a test result, a config export, a policy, a log sample. No artefact, no Met.
PartialRule: A control exists but evidence is missing or coverage is incomplete. Needs an owner and a fix date.
GapRule: No control. Needs an owner, a fix date and a compensating control if the system ships anyway.
Not applicableRule: Needs a one-line reason, for example "no retrieval, so LLM08 does not apply". Recheck at the next trigger.

Decide before the review which rows block release. A common rule is that a Gap on prompt injection when the system holds write-capable tools, on tenant isolation, or on human approval for irreversible actions stops the launch. Your own risk appetite and your customers' requirements set the real list. Write it down so the decision is not made under deadline pressure.

How do you reuse the review for a customer's security questionnaire?

The evidence column becomes your attachment list. Buyers' AI sections, including the SIG and the CSA AI-CAIQ, ask for the same facts: what data the model sees, who the providers are, how you test for injection, what a human approves, what you log. If your template is current, the answer is a link and a short paragraph. Our security questionnaire service covers how to answer when a row is a Gap: state the fact, the compensating control and a dated plan, rather than a yes you cannot prove.

Keep the template and attachments in one folder with a version history. When a buyer asks for evidence, you export, not rebuild. If a buyer wants a wider design review than this security view, the 30 questions in our AI architecture review cover cost, operations and technical debt as well.

What does this template not cover?

  • Privacy and data protection law. Under the DPDP Act, the core duties start on 13 May 2027, and what applies to your system is a question for counsel.
  • Certification. OWASP is a voluntary framework with no certificate. ISO/IEC 42001 certificates come from accredited certification bodies. Mono does not certify anyone.
  • Penetration testing. The template records test results. It does not replace a tester trying to break the system.
  • Model quality and fairness. LLM09 touches accuracy, but bias, explainability and harm to individuals need their own assessment.
  • Your contracts. Provider terms on training, retention and subprocessors need reading and filing next to the evidence.

Where Mono helps

Mono's cybersecurity team can run this review with your engineers: scope the system, test injection and tenant isolation, check tool permissions and build the evidence folder you attach to buyer reviews. The output is a dated evidence file and a list of fixes, not a certificate, and Mono is not a certification body or a law firm.

Sources: OWASP Top 10 for LLM Applications 2025 (genai.owasp.org/llm-top-10). OWASP Top 10 for Agentic Applications for 2026, published 9 December 2025 (genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026). Model Context Protocol specification 2026-07-28, Authorization and Tools (modelcontextprotocol.io). Digital Personal Data Protection Rules, 2025, G.S.R. 846(E). Last reviewed 20 November 2026.

FAQ

What is an AI security review template?

It is a table you fill in for each AI system, with one row per risk. Each row holds a review question, the control you run, the evidence that proves it and a dated status. Mapping the rows to the OWASP Top 10 for LLM Applications 2025 gives you a recognised structure. You re-run it when the model, data sources or tools change.

Can a company be OWASP compliant for LLM applications?

No. The OWASP Top 10 for LLM Applications is an awareness list and a source of guidance, and OWASP issues no certificate or compliance scheme. Say that your review is mapped to the OWASP LLM Top 10 and show the evidence for each item.

What evidence should I attach for prompt injection?

Attach a dated test set of direct and indirect injection attempts with the results, the list of untrusted content sources the model reads, and the permissions the model holds when fooled. Treat a pass as no failure found on that set on that date. Limiting the model's permissions is the stronger control.

Does the OWASP LLM Top 10 cover AI agents and MCP tools?

Partly. LLM06 Excessive Agency and LLM03 Supply Chain apply, but OWASP also published a separate Top 10 for Agentic Applications on 9 December 2025, with items such as Agent Goal Hijack and Identity and Privilege Abuse. For MCP tools, add rows for tool inventory, scopes, token handling, approval rules and server versions.

How often should I repeat an AI security review?

Repeat it whenever the model or provider changes, a new data source or tool is added, or a new customer type is onboarded, and at least once a year. Set the next trigger on the template itself so that stale evidence is noticed before a customer asks for it.

Get your AI system reviewed against the OWASP LLM Top 10.