AIM SHIELD Emergency
AI test report · Free

What do your AI's security tests say?

Upload the result of your tests and we explain it to you in Spanish: how many passed and how many failed in each category, which OWASP risks and articles of the AI Regulation they relate to and what to fix first.

We keep nothing and we launch no attacks. You run your tests; we only read the file you upload, in memory, and it is not recorded in any log.

Which files we read

We read the promptfoo JSON or JSONL (promptfoo eval -o, promptfoo redteam run -o) and the garak report.jsonl. Up to 5 MB. If the file is in another format, we tell you.

Giskard: its documentation says the scanner report can be exported to JSON, but it doesn't publish what that file looks like, and we don't guess. When it is published, we will add it.

A garak report.jsonl stores all the model's answers and can weigh many MB. To keep only the results per test, on your computer: jq -c 'select(.entry_type=="eval")' report.jsonl > resultados.jsonl

Your results

What we have seen

What to do first

    By category

      The report is downloaded from your browser. We do not keep a copy.

      How we read it

      One test, one result
      It passes, fails or gives an error. The ones that give an error are counted separately: you do not know whether they would have passed.
      By category
      We group by the type of attack you tested (instructions that trick, data that leaks, made-up answers…).
      OWASP and AI Regulation
      To each category we assign the closest OWASP risk and Articles 9, 15 or 55 of the AI Regulation. It is a guide, not an audit.
      What to do first
      We start with the categories with the most failures, with clear advice for each one.

      To find out whether your system is high risk, see Is my AI high risk?. To see which AI your team uses, the AI inventory. This is not legal advice.

      Frequently asked questions

      Do you launch attacks against my AI?

      No. You run your tests with your own tools and upload the results file to us. We only read it and explain it. We never connect to your system.

      Do you keep my file?

      No. The file is read instantly, in memory, and is not saved or written to any log: neither the text nor the names of your tests. If you close the page, it disappears. You download the printable report yourself, from your browser.

      Which files do you read?

      We read promptfoo's JSON or JSONL (promptfoo eval -o, promptfoo redteam run -o) and garak's report.jsonl. Up to 5 MB. If the file is in another format, we tell you: we do not guess.

      And Giskard's?

      Not yet. Giskard: its documentation says the scanner report can be exported to JSON, but it doesn't publish what that file looks like, and we don't guess. When it is published, we will add it.

      My garak report.jsonl is very large. What do I do?

      The report.jsonl keeps all the model's answers. A garak report.jsonl keeps all the model's answers and can weigh many MB. To keep only the results per test, on your computer: jq -c 'select(.entry_type=="eval")' report.jsonl > resultados.jsonl That way you only upload the result of each test.

      How do I count the tests?

      In promptfoo, each row in the file is a test: it passes or fails. Those that give an error are counted separately, because you don't know whether they would have passed. In garak, each test is a model answer assessed by a detector; answers that could not be scored are not counted.

      Why doesn't the category of my tests appear?

      In promptfoo, each test's category only appears in the files from security scans (promptfoo redteam run). In a normal evaluation we give you the totals and tell you so. In garak, the category comes from each probe's module; if it is not in our table, we count it as “other tests” and do not link it to OWASP.

      What does each OWASP code mean?

      These are OWASP's lists of the top ten risks for applications with language models (2025 edition) and for applications with agents (2026 edition). A specific test rarely fits a single risk perfectly: it is a guide to organise the work, not an OWASP audit.

      Do the articles of the AI Regulation oblige me?

      Articles 9 (risk management) and 15 (accuracy, robustness and cybersecurity) apply to high-risk AI systems, and Article 55 to general-purpose models with systemic risk. To find out whether yours is high-risk, see “Is my AI high-risk?” at /ia-alto-riesgo. If it is not, they work as good practice. This is not legal advice.

      If everything passes, is my AI safe?

      No. It means it has passed those tests. Other tests, other instructions or another model may give a different result. Repeat the tests every time you change something.

      Sources

      Sources consulted on 1 November 2026. This tool organises what your file says; it does not replace it or check it by itself, and a test passing does not prove that your system is secure.