Sample report
Sample report: what you receive
The report you receive, shown on two tests in our own lab: an AI email assistant built for the test, and DVNA, a public web app written to be vulnerable. This is what a penetration test report (VAPT report) from us contains: each finding says what an attacker can do, with the evidence and a fix.
Get the sample report
Leave your work email and the PDF link appears here.
Thank you. Download the sample report (PDF)
Summary
An AI email assistant attacked by emails from strangers
The same test assistant ran on five AI models. It received 50 attack emails and 10 normal ones. With no defense, 38 of 250 attack tries made it call its send-email tool to email the attacker from its owner's mailbox, and on one model 16 of its 50 did. The lab records each tool call and never runs it, so no email was sent. Three fixes, each tested on its own, brought the count to 0, 0 and 1 of 250. Every model still summarized all 10 normal emails under every fix.
An automated AI code review of a vulnerable web app
DVNA is a public web app written to be vulnerable, and its guide lists 19 flaws. We ran Claude Code's /security-review on it 3 times: it reported 10 of these flaws in every run and 12 in its best run, with no false alarm. Seven were never reported, among them forms that another website can submit in a logged-in user's name, and a session secret written in the code.
The 26 findings
Severity is ARELION's own judgment of what an attacker can do with each finding.
| ID | Finding | Severity |
|---|---|---|
| A-1 | An email from a stranger makes the assistant call its send-email tool from the owner's mailbox | Critical |
| A-2 | The assistant takes an order hidden in an email as an order from its owner | High |
| A-3 | With the mail tool removed, some replies still say an email was sent, or draft the attacker's email | Low |
| A-4 | A secret placed in the assistant's instructions appears in the model's reasoning text | Medium |
| A-5 | The screening call lets a few attack emails through | Medium |
| A-6 | The AI provider's own filter refused some requests | Info |
| B-01 | The automated review alone left 7 of the 19 known flaws unreported | High |
| B-02 | Command injection in the network test page (A1-CMDI) | Critical |
| B-03 | An uploaded file is turned back into running code (A8-DESER) | Critical |
| B-04 | The password reset token is the MD5 hash of the login name (A2-RESET) | Critical |
| B-05 | SQL injection in user search (A1-SQLI) | Critical |
| B-06 | The XML import reads files from the server (A4-XXE) | High |
| B-07 | Any user can change another user's password and details (A5-EDITUSER) | High |
| B-08 | The admin user list API is open to any logged-in user (A5-USERSAPI) | High |
| B-09 | The user list API sends password hashes (A3-HASH) | High |
| B-10 | A script can be stored in a product name (A7-SXSS) | High |
| B-11 | A script can be stored in a user name and runs in the user list (A7-DOMXSS) | High |
| B-12 | The calculator passes user input to an old math library with a known code execution flaw (A9-MATHJS) | Critical |
| B-13 | A script in a search link runs in the victim's browser (A7-RXSS) | Medium |
| B-14 | Forms that change data have no anti-forgery token (AX-CSRF) | Medium |
| B-15 | The session secret is written in the code (A2-SECRET) | Medium |
| B-16 | The app redirects to any address given in a link (AX-REDIRECT) | Low |
| B-17 | Every SQL query is written to the console log (A3-LOG) | Low |
| B-18 | The calculator shows a stack trace on bad input (A6-STACK) | Low |
| B-19 | The server announces its framework in a header (A6-XPB) | Info |
| B-20 | Logins, failed logins and other sensitive actions are not logged (A10-LOGGING) | Medium |
What is in the report
- A summary for management.
- Each finding with its severity, what an attacker can do, the evidence with its data file, and the fix.
- A coverage matrix against the OWASP Top 10 for LLM Applications 2025, the OWASP Top 10 for Agentic Applications 2026 and the OWASP web categories.
- The method in plain words, and the limits of the test.
A report on your system has the same parts. See the AI agent and chatbot security audit, the AI-built app security audit and the prices.