Skip to content

Meet DeepintShield at Bengaluru Tech Summit 2026, 17 to 19 November, BIEC Hall 1, booth H1A106. Book 20 minutes

Start For Free

Free tool · AI Runtime Risk Scan

Free AI security risk assessment: the AI Runtime Risk Scan

An AI security risk assessment checks how well your LLM apps and AI agents are protected in production. This free one from DeepintShield asks twelve yes, no or not sure questions across six areas, from prompts to audit evidence, and scores each area with a next step. No account needed.

  • Twelve questions, about two minutes
  • Scored in your browser
  • Mapped to the OWASP LLM Top 10
AI security risk assessment illustration: requests passing a pipeline of checkpoints with warning markers

How the scan works

Three steps, no account and no download. The scan runs on this page.

Answer twelve questions

Two questions for each of six areas. Answer yes, no or not sure.

See your score by area

Your browser works out a score for each area and one overall score.

Get a short list of fixes

Each area comes with one next step and the capability that covers it.

Start your AI security risk assessment

Choose Yes, Not sure or No for each question. You can go back at any time. An e-mail address is optional and only needed if you want the full report sent to you.

Area 1 of 6Input

Do you check every prompt for prompt injection and jailbreak attempts before it reaches the model?

Do you detect and mask personal data and secrets in prompts before they leave your network?

Area 2 of 6Retrieval

Do you scan retrieved documents for hidden instructions before they are passed to the model?

Can a retrieved chunk be blocked or redacted based on who is asking and what it contains?

Area 3 of 6Tool actions

Does every tool call an agent makes get a policy decision before it runs?

Do high-risk actions, such as payments or record changes, wait for a person to approve them?

Area 4 of 6MCP

Do you keep an allow-list of the MCP servers and tools your agents may use?

Would you notice if an approved MCP tool changed its description or behaviour after approval?

Area 5 of 6Output

Do you check model responses for personal data, secrets and unsafe content before users see them?

Do you check answers against your sources to catch hallucinations?

Area 6 of 6Evidence

Is every AI decision written to a log that cannot be changed without the change being detected?

Could you show an auditor which policy allowed or blocked a given AI action last month?

Risk scan questions

Do I need to sign up?

No. You see your score as soon as you answer the last question, with no account and no e-mail address. The e-mail field appears only after the score, for people who want the full report sent to them.

Do you store my answers?

No, your answers stay in your browser, where the score is worked out. If you ask for the full report, only your e-mail address and your area scores are sent, not the individual answers, so we can e-mail the report. See the privacy policy.

How is the score worked out?

Each answer earns points: Yes 2, Not sure 1 and No 0. Each area has two questions, so four points make 100%. The overall score is the average of the six areas: 75% or more is Strong, 40% to 74% is Partial and below 40% is Exposed.

Is this a full AI security audit?

No, it is a quick self-check, not an audit. For a deeper review, the Agent Risk Assessment pilot runs observe-only in your cluster for two weeks on one to three GenAI or agent apps and maps gaps against 25 controls. See pricing and pilots.

What should I do after I get my score?

Start with your lowest area. Each result links to the capability that closes that gap, such as Agentic Security for tool actions. To see how the controls fit together on one request path, read how DeepintShield works.

Want a walkthrough of your results?

Bring your score to a 20-minute call with an engineer. We will go through each area on your own stack.

Start For Free