A data-centre aisle between server racks

← All work

Three weeks in October.

A hacked website I was asked to rescue, an emergency PHP release in the middle of the night, a certificate authority that lost its keys, and the tool I built because of all three.

An alert at 01:55

PHP shipped an emergency release one evening in October. Two flaws, CVE-2026-109618 and CVE-2026-109619, let anyone run code on a server without logging in. The project's own notice said there were no complete mitigations and that updating was the only safe choice. The flaws had been found by researchers working with AI models.

I have a small tool that watches upstream builds. For each new build of the software I depend on, it works out what the build fixes, and a model running locally on my Mac classifies it. Security fixes raise an alert. At 01:55 the patched PHP images were built, and it woke me.

That night I patched the B2B AI platform I work on. By morning the hosting servers I look after were on the fixed release. A service still on PHP 7.4, a version that gets no fix, was moved to a current one the same morning, with the code changes it needed.

PHP2Shell · one night and one morning
  1. evening
    release

    emergency PHP release: CVE-2026-109618 and CVE-2026-109619

  2. 01:55
    alert

    patched images built; the build-watcher flags a security fix, classified by a local model

  3. night
    patch

    the B2B AI platform is on the fixed release

  4. morning
    patch

    the hosting servers I look after are on the fixed release

  5. morning
    upgrade

    a service on PHP 7.4, which gets no fix, moved to a current version

  6. morning
    lock down

    a legacy app that cannot upgrade yet: uploads refused, small bodies only, tight limits

  7. days later
    news

    exploitation confirmed in the wild

Times are from that night. Host names are left out on purpose.

One old application could not follow. Its framework stops at PHP 7.4 and a proper migration is days of work. So I cut down what could reach it: no file uploads outside the admin area, small request bodies only, tighter limits in PHP, workers recycled often. Then I put the migration on the list.

Exploitation in the wild was confirmed days later. By then every server I look after had been patched or locked down for days.

A rescue

In the same weeks I was asked to help with an optics clinic's website. It had been compromised for about a month through an old, known PHP flaw: rogue administrator accounts, a web shell hidden among stylesheets, and pages of spam that search engines had started to index.

I took forensic copies first, so nothing was lost before it was understood. Then I rebuilt the site from clean sources, removed the rogue accounts and rotated every password, key and database credential.

Then the hardening. The web server no longer runs PHP from upload folders or admin asset folders. The legacy remote-call endpoint is refused, exploit-shaped requests are refused, and logins are rate-limited. Administrator paths sit behind a one-time-passcode gate. PHP can reach the internet only through a proxy with an allow-list. A high school's website on the same server had its own problem, a fake plugin and tens of thousands of spam accounts, and got the same treatment.

rescue timeline · compromised website
  1. 01
    backup

    forensic copy of the infected files and database, sealed

  2. 02
    rebuild

    clean core, theme and plugins; rogue admins removed

  3. 03
    rotate

    passwords, salts, database credentials

  4. 04
    seal

    PHP in upload and admin dirs: 403 · legacy endpoint: 403 · exploit-shaped queries: 403

  5. 05
    egress

    PHP on an internal network, one allow-listed way out

  6. 06
    prove

    dropped test PHP file: alert in 15 s

Every step logged with where it lives and how to undo it.

Certificates I never chose to trust

In the same weeks a certificate authority lost the private keys to eight of its intermediate certificates. Two tools I rely on for private access, a tunnel client and a mesh VPN, shipped updates that stop trusting them.

I had never chosen that authority. My software trusted it by default. That is what a supply-chain problem looks like from below.

I swept every machine by hand and found what a hand sweep always finds: most hosts current, a couple behind, and one connector nobody remembered, months out of date.

The next day I wrote the design for an update checker. A day after that, tunnel-client and container updates were signed jobs with rollback. The first time one of them met something unexpected, it put the old version back by itself.

What came out of it: fleet-watch

For years my routine was a morning read: security news and advisories, then a decision about what to patch first. It works while the list is short. That month the list stopped being short.

The rescue showed the other limit. The site had been compromised through something old and known, and every step of the attack had left a trace that somebody could have seen at the time: a new file, a new administrator, a changed setting. A reading list cannot tell you what already happened on your own servers.

The first commit of fleet-watch is from the afternoon the rescued site went back online. It watches for those traces, checks what is installed against public vulnerability data, and rolls fixes out in stages. A PHP file dropped as a test raised an alert within 15 seconds. Today I run it on the servers I am responsible for, clients' included, and offer it as a managed service: fleet-watch.

Not sure what state your infrastructure is in?

I am happy to talk through how I would approach it.

Get in touch