What we attack

Services

Seven targets. For each one, an attack path as we would write it in a report, and what we do not touch.

The scenarios are fictional. They name classes of flaws and what they mean for the business, not exploitation methods.

The exact scope is set on the scoping call, in writing, before the first test.

01 of 07Web applications and portals

Web

Another customer's account.

Fictional scenario. Customer portal of an invoicing software vendor. Two test accounts: an owner and a plain reader.

Attack path

  1. We replay every action of the interface with the reader account. The screen hides the export button, but the route behind it still answers.

  2. Broken function-level access control: the team CSV export checks the session, not the role. The reader downloads the billing contacts of the whole organisation.

  3. In that file, a "company name" field that any member can edit. The support tool displays it back without neutralising it: stored XSS.

  4. When an agent opens the record, the script runs with their rights. We stop at proof: a capture of the admin console, opened with an agent's session.

Where it leads

A plain reader becomes, through a support agent, administrator of every customer account.

  • OWASP A01: broken access control
  • Stored XSS
  • Chained flaws

Out of scope

  • Code review: without the source code, we test from the outside
  • Third-party applications you do not administer

02 of 07REST, GraphQL, tokens

API

Change a number in the URL.

Fictional scenario. Billing API of a B2B SaaS, documented in OpenAPI, called by the web app and the mobile app.

Attack path

  1. We start from the documentation and the mobile app traffic to list the routes, including a /v1 version the front end no longer calls but the server still serves.

  2. IDOR on the billing API: with account A's token, the invoice read route accepts the identifier of an invoice owned by account B. Identifiers are sequential.

  3. The profile also accepts fields the form never sends. Adding the role field to the update, the server writes it: mass assignment.

  4. The test account becomes administrator of its own organisation. The organisation list filters nothing: we can see the others, and we stop there.

Where it leads

Every customer's invoices can be read, then accounts taken over: addresses, amounts, contacts.

  • API1: object-level authorization (BOLA)
  • API3: object properties (BOPLA)
  • API9: version inventory

Out of scope

  • Load or saturation testing
  • A third-party provider's API, outside your control

03 of 07iOS, Android and their API

Mobile

What the app keeps on the phone.

Fictional scenario. Delivery app with customer and courier accounts, built-in payment.

Attack path

  1. We install the app on a test device and read what it stores: local database, preferences, logs, system backup.

  2. In the package, the key of a third-party SMS service, in clear text. It lets anyone send messages at your expense, under your sender name.

  3. A deep link opens the e-mail change screen without asking for the password again. The app trusts the link, not the user.

  4. A link sent to a customer changes their account e-mail. The password reset then lands with us.

Where it leads

Taking over a customer account in one click, without knowing the password.

  • MASVS-STORAGE
  • MASVS-PLATFORM: deep links
  • Embedded secrets

Out of scope

  • Jailbreak or root as an attack prerequisite, unless agreed beforehand
  • App stores and operating system services

04 of 07What your configuration exposes

Cloud

A subdomain nobody watches any more.

Fictional scenario. A company hosting its product and its backups with a public cloud provider.

Attack path

  1. Passive reconnaissance: public certificates, DNS, repository names. A test subdomain still points to a service you cancelled.

  2. Subdomain takeover: we claim the name with the provider and the page answers under your domain, with access to cookies set for the parent domain.

  3. A backup container allows anonymous listing and reading. We count the files, open one, and stop.

  4. A server-side URL preview feature can reach the machine's metadata address: SSRF to the instance service. We read the name of the attached role, not its tokens. It is far too broad.

Where it leads

Production backups can be read, and there is a path to the machine's temporary credentials.

  • Subdomain takeover
  • Public storage
  • SSRF to metadata

Out of scope

  • Configuration review from inside the cloud account
  • Any test against the provider's own infrastructure

05 of 07What answers on the Internet

Network

The front door nobody remembers.

Fictional scenario. Industrial SME: a head office, two sites, remote access by VPN, in-house mail server.

Attack path

  1. We list the organisation's addresses and what answers on them: VPN gateway, mail server, admin interface, a device forgotten in a rack.

  2. The remote access gateway reports a version hit by a public pre-authentication code execution vulnerability. We confirm it from the version and the service's behaviour, without running code on the device.

  3. A UPS admin interface answers with the manufacturer's default credentials.

  4. That is a foot in the network, or proof that there is one. The internal network is out of the mandate: the test stops at the border.

Where it leads

A direct way inside, without a valid password and without any employee clicking anything.

  • Exposed vulnerable service
  • Default credentials
  • Admin interface

Out of scope

  • Internal network, unless under a separate mandate
  • Addresses whose control you cannot prove
  • Any denial-of-service technique

06 of 07OAuth 2.0, OpenID Connect, SAML

Identity

Walk in without a password.

Fictional scenario. SaaS product with sign-in through a corporate identity provider and a social account.

Attack path

  1. We map the authorisation flow: client identifier, registered redirect addresses, state value, requested scopes.

  2. The server validates the redirect address by prefix instead of requiring strict equality. Combined with an open redirect on the site, it sends the authorisation code to a domain we control.

  3. Account linking uses the e-mail address announced by the provider, without requiring it to be verified. We create a social account with someone else's address.

  4. On sign-in, the product attaches our social identity to the target's existing account.

Where it leads

Taking over the account of any user whose e-mail address we know, without touching their password.

  • redirect_uri badly validated
  • Account linking
  • Account takeover

Out of scope

  • The identity provider itself, if run by a third party without its authorisation
  • Phishing of users

07 of 07Assistants, agents, retrieval-augmented search

LLM

The model obeys what it reads.

Fictional scenario. Customer support assistant plugged into the help base, the order history and a refund tool.

Attack path

  1. Indirect prompt injection: a ticket contains instructions. The assistant reads them with the same weight as yours.

  2. The "look up an order" tool takes a number without checking who owns it. The model is the only safeguard, and it can be talked round: excessive agency.

  3. The reply is rendered as Markdown. A remote image in the output triggers a request to a third-party domain, with the data in the URL: improper output handling.

  4. The system prompt, repeated word for word, lists the internal tools and their parameters.

Where it leads

Other customers' orders can be read, without an account, from a plain ticket.

  • LLM01: prompt injection
  • LLM06: excessive agency
  • LLM05: output handling

Out of scope

  • The provider's model itself: bias, generated content, alignment
  • Evaluating the quality of answers

The limits

What we do not do.

These limits are fixed. They protect your systems and our legal footing.

  1. Denial of service or load testing
  2. Phishing and social engineering, unless an explicit contract provides for them
  3. Physical intrusion
  4. Testing anything without a written authorisation signed by an empowered person
  5. Accessing more data than the minimal proof needed to demonstrate a flaw

Who it is for

It suits you if

  • You want to pay only for confirmed vulnerabilities, with a mission cap (30 days, adjustable) set by contract
  • Your scope is defined and you can prove you control it
  • You are looking for test evidence to feed an ISO 27001, SOC 2, NIS2 or DORA programme
  • You have a product in production and someone to triage reports within 5 business days

Is your target on the list?

Leave your site and email: we get back to you within 2 business days.