# AI security check

Ten questions about how you use AI. You immediately see a score per area and the three measures that pay off most. It takes about three minutes.

Your answers stay in your browser and are not stored or sent anywhere. This is a first estimate, not an audit.

The ten questions follow the two current OWASP lists for AI security. Each measure states which risk it addresses. OWASP Top 10 for LLM Applications (2025): https://genai.owasp.org/llm-top-10/ · OWASP Top 10 for Agentic Applications for 2026: https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/

## 1. Do you know in which region your prompts and completions are processed?

- Yes, contractually agreed and technically enforced
- Yes, based on what the vendor says
- No, or I do not know

**Enforce the processing region technically.** Route all model requests through a gateway that only allows European endpoints and refuses requests for which that is not possible.

## 2. Is personal data masked before it goes to a model?

- Yes, automatically
- Sometimes, manually by staff
- No, or I do not know

**Mask personal data automatically.** Replace structured data deterministically and run free text through PII detection, so it does not depend on staff paying attention.

## 3. Do staff use public AI tools with company or customer data?

- No, there is an AI environment of our own that they use instead
- There is a policy, but no control
- Yes, or I do not know

**Give staff an AI environment of their own.** A ban only works with a good alternative. Offer your own chat environment that is just as easy and where you decide where data goes.

## 4. How is your AI environment or AI application reached?

- Only after signing in with single sign-on
- With its own password, reachable from the internet
- I do not know

**Put identity checks in front of the application.** Let nobody reach the application without first signing in through your identity provider, and also check that admin screens are not directly on the internet.

## 5. How does your AI handle outside content such as emails, documents and web pages?

- As untrusted and separated from the instruction, or we process no external content
- The same as our own instructions
- I do not know

**Treat external content as data.** Mark everything that comes from outside as untrusted and keep it separate from the instruction, so hidden instructions in a document are not read as a command.

## 6. Do you test your AI application for prompt injection?

- Yes, with every release
- Once, at delivery
- No

**Make attack tests part of every release.** Record a set of known attacks and run it automatically on every change, just like ordinary tests.

## 7. What permissions does an AI agent or AI integration have in your systems?

- Only what the task needs, or we have no agents
- The same as the user or a broad service account
- I do not know

**Give agents least privilege.** Give every agent its own account with only the permissions for its task. What an agent may not do, an attacker cannot make it do either.

## 8. Is human approval required for risky actions such as deleting, paying or deploying?

- Yes, always
- For some actions
- No

**Have a human approve risky actions.** Decide which actions are irreversible or costly and let the AI only prepare them. A human carries them out or approves them.

## 9. Can you reconstruct afterwards what an AI application or agent did?

- Yes, there is an audit log per action
- Only through general application logs
- No

**Record every action in an audit log.** Record per action who or what performed it, with which input and with what result, in a place the agent itself cannot change.

## 10. Do you measure the quality of AI answers on every change?

- Yes, with a fixed test set
- Occasionally, by hand
- No

**Measure quality with a fixed test set.** Compile a set of real questions with expected answers and measure every version against it, so you see whether a change improves or degrades.

## Pick an attack and see where it stops

- **Scanner hunts a known vulnerability.** As soon as a flaw in widely used software becomes public, bots scan the whole internet within hours for servers that still have it. With Cloudflare Zero Trust the server opens an outbound tunnel itself and no port is open; a request without a valid identity never reaches the application, even when it has not been patched yet. Patching is still needed, but it buys you time.
- **Stolen password.** Passwords leak through phishing and reuse. Behind Zero Trust a password alone is not enough: sign-in goes through your identity provider with a second factor, per application, and access can be revoked at once when someone leaves.
- **Hidden instruction in a document.** An attacker hides a command in a PDF or email that an employee asks the AI to summarise. Because external content stays separated from the instruction, the text is summarised and not carried out. This layer is not watertight, which is why two more follow.
- **Manipulated agent looks for secrets.** Suppose an injection does succeed and the agent is told to look up passwords or keys. The agent only has permissions for its task and cannot reach them.
- **Agent wants to deploy or delete something.** An agent, manipulated or simply wrong, prepares an irreversible action. It is only carried out after a human has looked at it and approved it.

---
Source: https://mg-devs.nl/en/ai-security-check/
