~/blog/the-server-next-doorPOSTED
by Shady Nathan Tawfik · October 29, 2026 · 6 min

The Server Next Door: When the Website Is the Cleanest Part of the Estate

A pattern essay about a medical practice whose website got everything right - no third parties, no cookies, no forms, nothing for a consent banner to gate - and whose real risk sat two hosts away: a mail server twelve months past its last security update, answering on any host name and writing it into the login redirect, a legacy gateway with a certificate that expired in 2023, a package manager on the open internet, and a privacy notice describing a form that does not exist.

Interactive[blog-2.0]

TL;DR: This was the cleanest website we reviewed this year: 242 resource tags across nine pages, zero external hosts, zero cookies, zero storage keys, zero forms — and therefore no consent banner, correctly. The risk was two hosts to the side. A mail server past its end of support and eight updates behind, publishing its build and its internal server name to anonymous callers, accepting NTLM, and answering on any host name — which turns it into a phishing platform that serves the genuine login mask under a lookalike domain with a valid certificate. A legacy gateway with a certificate that expired in 2023 and a PPTP port still open. A package manager reachable from the open internet, with the login form served over plain HTTP. And a privacy notice that describes a contact form that does not exist, while the site invites patients of a medical practice to send health data by unencrypted e-mail.

The scenario

A medical practice asked me to look at its website. Several independent practices share one building, one website and one mail system — a structure that is common in German healthcare and that quietly multiplies every incident by the number of practices involved.

The website itself is genuinely well built. No Google Fonts. No analytics. No pixel. No map embed — just a link. No cookie banner, because there is nothing to gate. No forms at all, so no CSRF token, no upload field, no reflected parameter to worry about. The install tool is locked, the maintenance endpoints redirect to the login, and a foreign Host header never reaches the site's virtual host.

If the engagement had stopped there, the report would have been one page of praise. It did not, because a website is not a perimeter.

What "end of support" actually means

The first neighbouring host was the mail server. It answered with a build number in a header, on every anonymous request, on three different host names:

~/code/text text
HTTP/2 302
x-owa-version: 15.1.2507.61
x-feserver: EX01

That build is the October 2025 security update of a product whose support ended on 14 October 2026. The current build of the same line is .77, released four days before the review. So the installation is twelve months of security updates behind, on a product that will never receive another one, and it announces the gap to anyone who asks.

Exchange is not a peripheral. For a practice it holds every appointment request, every finding, every invoice conversation. The update window that was skipped contains — among others — a cross-mailbox read class and an OWA-path remote-code-execution class. Neither needs to be exploited cleverly. They need one valid password: from an old dump, from a reused credential, from a mail that looked like it came from a colleague.

And the server tells you its internal name while you are at it. EX01 is not a secret in a hardened design, but publishing it to anonymous callers removes the reconnaissance step entirely.

The part that turns a mail server into a phishing platform

Then we sent a request with a host name that does not exist:

~/code/text text
$ curl -sk --resolve probe:443:<ip> -H "Host: evil-example.test" https://probe/owa/
HTTP/2 302
location: https://evil-example.test/owa/auth/logon.aspx?url=https%3a%2f%2fevil-example.test%2fowa%2f&reason=0

The server accepted the name and wrote it into the login redirect. Which means: register a lookalike domain, point it at the same IP, issue a free certificate, and the genuine login form appears under that name with a valid padlock. No clone, no pixel-perfect rebuild, no hosting of your own. The most credible phishing sender is the real server wearing a different name.

For a practice, the recipient list is already public: the team page, the professional registers, the health-network addresses in the contact page. Phishing does not fail because the target is careful. It succeeds because the address is right, the name is right and the login page is real.

The quiet path: plain HTTP and an open package manager

Two more findings belong together, because they are the quietest way in.

The homepage redirects to HTTPS. The login does not: http://…/contao/login returns the full form with status 200, and because the form has no action, the POST goes to the same plain-HTTP URL. The package manager — a Composer console with write access to the filesystem — behaves the same way, and its HTTPS response carries no HSTS, so a browser never pins it.

Nobody has to attack a server. Someone on the same Wi-Fi reads the credentials and is then legitimately inside the back end. From there the manager installs a package with arbitrary code: database, files, backups. The website was the entrance, never the target.

The documentation as a finding

The last group of findings needed no technical access at all — only reading.

  • The contact page invites patients to send appointment requests by e-mail. For a medical practice, that content is special-category data.
  • The privacy notice cites Art. 6 only. Art. 9 appears six times, all inside boilerplate sections about data-subject rights. There is no legal basis for processing health data and no instruction about a secure channel.
  • The same notice describes a contact form, four times, including a consent captured on submit. On the site there is no <form> element at all. The consent could never have been given.
  • Hosting and mail operators are not named, although their processing is described precisely.
  • Three pages name three different controllers — the privacy notice, the imprint and the team page disagree about who is responsible.

That last one is the one that costs money, and not because of a hacker. An access or erasure request lands with a body that is not the controller according to its own statement. The one-month deadline cannot be met cleanly. Proceedings in data protection rarely start with a breach; they start with an answer that came too late.

What this changes about reviewing a web estate

Three habits came out of this engagement.

Review the estate, not the URL. A clean front end is a good sign and a bad proxy. The perimeter is every host that shares the domain, the certificate and the brand — and every provider that can change them.

Treat end of support as a harder finding than any CVE count. A missing patch can be applied tomorrow. A product without support will never be fixed, and the organisation that runs it has usually not made that decision consciously.

Read the legal pages as technical evidence. A privacy notice that describes a form that does not exist proves template reuse without verification. A notice that cites Art. 6 in a medical context proves the data class was never modelled. Both are findings, and both are provable without touching the system.

Boundary

Everything here was collected from public sources: headers, DNS, certificate transparency, published release lists, and the site's own pages. No login, no payload, no exploitation, no data copy. The attack chains are described from version, configuration and advisory data — not executed.

Remediation is not part of this essay, and not part of the audit it came from. Hardening is a separate engagement.

Shady Nathan Tawfik

Founder & Cyber IT Security Engineer in Bielefeld. Offensive security, audits and full-stack builds. More at

For educational purposes only. Every reproducible command targets my own lab environment (localhost), never a live system. © 2026