dlvx.co Board of Innovation

Setup guide · prepared by DLVX · 8 September 2026

Running your own AI security agent on boardofinnovation.com

For Nico, or whoever holds shell access to the Hoasted account. Nothing in this document changes your site by itself.

Why this exists

On 18 August 2026 boardofinnovation.com was actively serving a hidden Dutch casino backlink block (Duckysino, linking to mijnmarbus.nl) on every page, planted through a fake plugin disguised as a caching tool. It was found, traced, and cleaned. The fake plugin and a second, empty backdoor file were archived outside the web root with verified hashes so the evidence is not lost.

Since then there has been one script, boi-security-check.sh, already installed and authorised on both Hoasted accounts. It reads the server the way an attacker would and prints PASS, WARN, or FAIL for a fixed list of checks. It found real, current problems on the last run, 3 September 2026: a WordPress account named admin on both sites (published openly on philippederidder.com), WordPress core and nine plugins behind on boardofinnovation.com, Elementor two major versions behind on philippederidder.com, and a world writable folder plus an exposed error_log sitting inside the public web root.

That script is solid, but it is a report. Somebody still has to read it every month, decide what matters, and go fix it by hand. This document sets up something more useful: Claude Code, running with the same read access the script already has, so it can run the check, explain the result in plain language, and, once you approve, make the fix itself, check its own work with real screenshots, and confirm nothing on the live site broke.

Nothing here removes the script. It still runs first, every time. Claude Code reads its output, not the raw server, wherever that is enough.

What you need before you start

Step 1: install Claude Code

npm install -g @anthropic-ai/claude-code

Then start it once to sign in:

claude

It opens a browser tab to log in with an Anthropic account, or a Claude subscription if you already have one. Once signed in, that machine stays authenticated.

Step 2: give it its own SSH key, not yours

Do not point Claude Code at your personal SSH key. If it is ever compromised, or you want to revoke its access, you do not want to also lock yourself out. Generate a separate key just for this:

ssh-keygen -t ed25519 -f ~/.ssh/boi-claude-agent -C "boi-security-agent"

That creates two files: boi-claude-agent (private, stays on this machine only, never shared, never pasted anywhere) and boi-claude-agent.pub (public, safe to share, this is what gets installed on the server).

Print the public key so you can copy it:

cat ~/.ssh/boi-claude-agent.pub

Log into cPanel on the Hoasted account (or SSH in with your own existing key) and add that output under SSH Access → Manage SSH Keys → Import Key, or append it directly to ~/.ssh/authorized_keys on the account.

Then add a shortcut so Claude Code, and you, can connect without retyping the details. In ~/.ssh/config on your machine:

Host boi-hoasted
  HostName s1151.hostingsecure.com
  User piexels
  Port 22
  IdentityFile ~/.ssh/boi-claude-agent

Test it:

ssh boi-hoasted "echo connected"

If that prints connected, the key works and Claude Code can use the same boi-hoasted shortcut.

If DLVX should also have this access If you want us to be able to run the same checks or help with a fix, generate a second key pair the same way, but send us only the hostname, port, and account username (the same lines as the Host boi-hoasted block above, minus the key itself), and confirm you have added the public key we send you to that account's authorized_keys.
Never send a private key A private key is the only secret in this whole setup, and it should never leave the machine it was generated on, over WhatsApp, email, or any chat, to us or to anyone. If one is ever pasted into a conversation by mistake, treat it as burned and generate a new pair.

Step 3: the prompt

Paste this whole block as the first message in a new Claude Code session, run from the machine with the SSH key set up above. Claude Code will ask before it does anything that changes the live site.

You are running a security maintenance pass on a WordPress site,
boardofinnovation.com, hosted on Hoasted, reachable over SSH at the host
alias "boi-hoasted". A hidden casino backlink injection was found and
cleaned here on 18 August 2026, via a fake plugin disguised as a caching
tool, so treat any unfamiliar plugin, mu-plugin, or theme file as
suspicious until proven otherwise.

Work in this order, every time:

1. Read first. Run the existing script over SSH before anything else:
   ssh boi-hoasted "./boi-security-check.sh --out security-report.txt"
   Read the output. Do not skip this even if you think you already know
   the state of the site.

2. Explain the result in plain language before touching anything. For
   each FAIL and WARN, say what it means and what you propose to do
   about it. Wait for my explicit go-ahead on each fix category before
   acting. Never bundle an unrelated change into an approved fix.

3. Follow WordPress security best practice for every fix:
   - Take a real backup (JetBackup in cPanel, or ask me to) before any
     change that touches files, plugins, themes, or the database.
     Confirm the backup finished before proceeding.
   - Prefer the smallest fix that solves the actual finding. Do not
     "clean up" unrelated files, plugins, or settings while you are in
     there.
   - Update WordPress core, plugins, and themes only from the Dashboard
     or WP-CLI, never by manually overwriting files, so version
     tracking stays intact.
   - Never rename or delete a file you believe is part of an active
     compromise. Move it outside the web root and record its hash and
     original path instead, the same way the August incident was
     contained. If you find anything that looks like the August
     pattern again, stop, do not clean it up yourself, and tell me
     immediately.
   - Fix file and folder permissions to 644 and 755 respectively,
     never wider, and never leave anything world-writable.
   - Never leave a backup, archive, or .env-style config file inside
     public_html or any other web-servable folder.
   - Do not leave an administrator account named something guessable
     like admin. If one exists, create a properly named replacement,
     move ownership of content across, and only then remove the old
     account, with my confirmation before the removal step.
   - Turn on login rate limiting if the report shows active guessing
     attempts, and flag two factor authentication as a recommendation
     if it is not already on.

4. Before making any change that could affect what a visitor actually
   sees (a plugin update, a theme change, a setting that touches
   layout, fonts, forms, or navigation), take a screenshot of the real
   homepage and at least 3 other real pages first (pick pages that
   exercise navigation, a form, and a content-heavy page). Save them
   clearly labelled as "before".

5. Make the one approved change.

6. Take the same screenshots again, labelled "after", and compare them
   side by side. Confirm nothing moved, broke, disappeared, or changed
   appearance in a way that was not the intended fix. If anything looks
   wrong, say so plainly and offer to roll back from the backup taken
   in step 3, rather than trying to patch it further.

7. Confirm nothing was lost. Re-run the security check script from
   step 1 and confirm the specific FAIL you were fixing is now gone
   and nothing else newly failed. Separately, load the homepage and
   the page you changed with a real HTTP request and confirm both
   still return 200, and if there is a contact or lead form on the
   affected page, confirm the form still renders (do not actually
   submit it unless I ask you to).

8. Report back in plain language: what you found, what you changed,
   what you verified, and what is still open. Do not say something is
   fixed unless you have the before/after screenshots and the re-run
   script output to show for it.

If anything in the report looks like a live, active attack rather than
a routine hygiene finding, stop everything else and tell me first
before doing any cleanup at all.

Step 4: run it whenever you want a check

Start Claude Code in a terminal on the machine with the key set up, paste the prompt above once, and from then on you can just ask it things like "run the monthly check" or "is anything urgent right now" in the same session, and it will follow the same read-first, screenshot, verify pattern every time.

Optional: run it automatically every day

The script itself is safe to run unattended, since it only reads. Point it at a fixed time every morning with a cron job, so a fresh report is always waiting rather than depending on someone remembering to run it.

Log into the Hoasted account over SSH and run:

crontab -e

Add this line to run the check every day at 8:00am server time:

0 8 * * * cd ~ && ./boi-security-check.sh --out security-report-$(date +\%Y-\%m-\%d).txt

That writes a dated report file every morning and changes nothing else. It does not run Claude Code itself, since Claude Code needs an active session and a human to approve fixes, which is the correct amount of caution for anything that can change the live site. What it does give you is a standing daily paper trail: if something breaks, you can look at yesterday's and today's report and see exactly when it changed, rather than only finding out at the next monthly check.

If you want Claude Code itself to read that daily file and flag anything new without a human starting the session, that is possible too, but it means giving Claude Code's key the ability to send you a message unattended (email or a webhook), which is a separate decision with its own tradeoffs. Worth a conversation before setting it up, not something to default into.

What this does not do

Same limits as the underlying script: no database scan, no restore test, no email deliverability check, and no comparison of premium or custom code against a vulnerability database, since those need a human judgement call at least once or twice a year. Claude Code following this prompt makes the routine 90% faster and leaves a clear evidence trail for the rest; it does not replace an actual incident responder if something like August happens again.

DLVX · prepared for Board of Innovation · 8 September 2026