Skip to content

How we test

Every Grapefruit penetration test follows the same path: map your attack surface, attack the web app, APIs and AI features together, prove each finding with a captured request and response, rank it, and re-test your fix. Built for teams that ship AI.

How does a Grapefruit penetration test run?

A Grapefruit test runs in six phases: kickoff, mapping, attack, proof, reporting and re-test. You add a hostname, publish one DNS TXT record and sign online. Testing starts, and results land in your portal within hours of kickoff. Every phase hands you something you can use.

PhaseWhat happensWhat you receive
KickoffYou add your hostname, publish one DNS TXT record and sign the testing authorization onlineA verified domain and a test that starts today
MappingPages, endpoints, roles, LLM features and agent tools are discovered and chartedA test that covers your whole product, including the AI
AttackWeb app, APIs, LLM features and agents are attacked together with real exploitationFindings that follow an attack across layers
ProofEach successful attack is captured as a request and responseEvidence your engineer can paste and replay
ReportingFindings are ranked critical, high, medium or lowAn online portal and a downloadable report
Re-testYour fixes are attacked againA dated record of what is new and what got fixed

How do we map your attack surface?

We start by charting everything an attacker could touch: pages, forms, REST and GraphQL endpoints, user roles, chat and RAG features, and the tools your agents can call. Using the test accounts you supply, the mapping goes behind the login, where most of the valuable functionality lives.

The map decides the quality of the test. An endpoint nobody listed is an endpoint nobody attacked. So the test finds its own targets, and it records what each role can see so the roles can be played against each other.

How do we attack app, API and AI together?

The web app, the APIs, the LLM features and the agents are attacked in one test, because real attacks chain across them. A weakness in the chat feature becomes the way into the API. A weakness in the API becomes the way to another customer's data. One test follows the whole path.

A typical chain: a document uploaded by one user carries hidden instructions. The assistant reads it, calls an internal endpoint and returns another tenant's invoices. You see every step, in order, in one place.

  • Web app: authentication, session handling, access control between roles, injection and business logic.
  • API: object and function level authorization, authentication, property exposure and server side request forgery.
  • LLM features: prompt injection, sensitive information disclosure, system prompt leakage and improper output handling.
  • Agents: goal hijack, tool misuse, privilege abuse and memory poisoning.

How is each finding proven?

Every finding arrives with the exact request and response that proves it. Your engineer pastes the request, runs it and watches the attack work. There is nothing to interpret and nothing to take on trust. After the fix, they run the same request again and watch it fail.

Proof is what turns a report into a fix. Your engineer skips the hours of reproduction and goes straight to the code. Your security lead can show any stakeholder the attack working, live.

How are findings ranked and delivered?

Findings are ranked critical, high, medium or low and delivered in an online portal, each with its captured request and response. You see the most serious issues first, with what was attacked, what broke and the evidence. A downloadable report packages the same findings for auditors, customers and your board.

  • Severity first: critical and high findings sit at the top, so your first hour goes to what matters.
  • Evidence attached: the request and response live with the finding.
  • One place: app, API and AI findings sit in the same list, so a chained attack reads as one story.
  • Downloadable report: one document for your auditor, your customer's security team or your SOC 2 evidence folder.

What happens after you fix something?

Fix something and it is re-tested. The attack that proved the finding is run against your fix, so you know the hole is closed. On a maintenance plan the re-test happens on your next scheduled run, weekly or monthly, with nothing to book and nothing extra to arrange.

Each run is dated and kept in the portal. It shows what is new and what got fixed since the last run. Over a few months that becomes a ready evidence trail for a SOC 2 Type II audit window.

Which standards does the testing follow?

Grapefruit testing follows the OWASP Web Security Testing Guide (WSTG) v4.2 for web apps, the OWASP API Security Top 10 2023 for APIs, the OWASP Top 10 for LLM Applications 2025 for LLM features, the OWASP Top 10 for Agentic Applications for 2026 for agents, and MITRE ATLAS for adversary techniques against AI.

These are the references auditors, customers and security teams already use. When a buyer's questionnaire asks which methodology your pentest followed, the answer is on this page.

How do you start a test?

Start today. Add your hostname, publish one DNS TXT record and sign the testing authorization online, which takes a few minutes. Choose a full penetration test at a fixed quote, or a maintenance plan with public prices and card checkout. There is no demo call. See pricing, or read the sample report first.

Questions

How long does a Grapefruit penetration test take?
Results land in your portal within hours of kickoff. Kickoff itself takes a few minutes: add your hostname, publish one DNS TXT record and sign the testing authorization online. There is no scheduling queue and no demo call between you and your first finding.
Do you test behind the login?
Yes. You supply test accounts and the test runs authenticated, as each role you provide. That is where access control flaws live, such as one customer reading another customer's data, so authenticated testing covers the functionality that matters most to your users.
What does a finding look like?
Each finding has a severity of critical, high, medium or low, a description of what was attacked and what broke, and the captured request and response that proves it. Your engineer can paste the request, run it and reproduce the attack. The sample report shows a full example.
Are AI features tested separately from the web app?
No. Your web app, REST and GraphQL APIs, LLM features such as chat and RAG, and AI agents with tools are attacked in the same test and reported in one place. Attacks that start in the AI and finish in the API are followed end to end.
Do you re-test after we fix a finding?
Yes. Fix something and it is re-tested. Subscribers on a maintenance plan get the re-test on their next scheduled weekly or monthly run, and the portal records the date along with what is new and what got fixed since the previous run.

Standards and sources