Ethernet cables plugged into a network switch

Security

The bugs are being found by machines now. The patching is still done by people.

This is what one autumn of AI-found vulnerabilities looked like from the inside, and what it means for your servers, whether you run twenty or twenty thousand.

The shift

> 100,222
CVEs in 2026 by 13 Nov, against 48,185 in 2025

Last year about 48,000 vulnerabilities were published, a record. This year passed 100,000 by mid-November. One of the people who counts them says that hunting for vulnerabilities with AI is driving most of what they see.

−7 days
average time to exploitation: the attack comes before the patch
AttackPatch−7 days

Speed changed too. Google's incident responders now put the average time to exploitation at minus seven days: on average, the attack comes a week before the patch. Nearly three in ten exploited vulnerabilities last year were already in use on or before the day they were published.

Sources: The Register / VulnCheck (2026) (opens in a new tab), The Stack (2025) (opens in a new tab), Google M-Trends, VulnCheck State of Exploitation. CVE counts as of 13 Nov 2026.

The skill barrier is gone. Attacks that once needed an expert now need a gaming PC and an uncensored model.

An emergency release, and an alert at 01:55

PHP2Shell was two flaws in PHP, CVE-2026-109618 and CVE-2026-109619, that let anyone run code on a server without logging in. They were found by researchers working with AI models. The fix shipped one evening in October, with a notice that updating was the only safe choice.

My build-watcher woke me at 01:55, when the patched PHP images were built: it reads what each new upstream build fixes, and a model running locally on my Mac classifies it.

That night I patched the B2B AI platform I work on.

By morning the hosting servers I look after were fixed.

A PHP 7.4 service was upgraded, and one legacy application that could not upgrade was locked down instead.

Days later, attacks in the wild were confirmed. By then every server I look after had been patched or locked down for days. The whole night, hour by hour: Three weeks in October.

A site compromised for a month, rebuilt and sealed

In the same weeks I was asked to help with an optics clinic's website that had been compromised for about a month. I took forensic copies, rebuilt it from clean sources, and sealed the server: PHP only where PHP belongs, administrator paths behind a one-time passcode, outbound traffic through an allow-list. The full story.

Certificates you never chose to trust

A certificate authority lost the private keys to eight of its intermediate certificates. A tunnel client and a mesh VPN I rely on shipped updates that stop trusting them. I had never chosen that authority; my software trusted it by default.

A sweep by hand found most machines current, a couple behind, and one forgotten connector months out of date. So I stopped sweeping by hand: within two days those updates were signed jobs with rollback, the start of the update side of fleet-watch.

Agents need access. So the access has to be narrow.

In July, AI models under evaluation at OpenAI escaped their sandbox. Their only route to the network was a package-registry proxy. They found a flaw in it that nobody had published, went through, and ended up inside Hugging Face's systems. Both companies disclosed it. The sandbox was a light one, of the kind many coding agents use by default.

I run AI agents every day, on real systems. An agent only does useful work when it has real access. So the question is never whether to give access. It is how narrow, how short-lived, how well logged, and who approves. After July I went well beyond a light sandbox:

Egress through an allow-list

Agents reach the network only through a proxy with a list of permitted destinations. Everything else is refused and logged.

SSH certificates per task

Access to a server is a certificate minted for one task, restricted to one source address and short-lived.

SQL through a controlled integration

Database queries go through a controlled integration with its own limits. The agent never holds a database password.

A secrets runner

A command that needs a secret gets it injected at run time. The value is scrubbed from everything the agent can read.

A guard the agent cannot edit

The rules live in a guard owned by root, outside the agent's reach. It cannot rewrite its own limits.

A person approves

Risky actions stop at a deterministic approval step: the same request always asks the same question, and a person answers it.

I set this up for teams as part of AI-assisted engineering.

A login bug found by agent swarms

In November a researcher, Ozgun Kultekin, published a flaw in OpenSSH on systems built with the musl C library and no PAM: for some accounts, an overlong password was accepted. He found it by sending agent swarms at many builds of the server, about 17,600 unusual login attempts in all. It is CVE-2026-123158, fixed in OpenSSH 10.6p2, and it was on the US list of exploited vulnerabilities four days after publication.

It affects small Linux builds, such as Alpine and router firmware, under specific conditions. Mainstream server distributions are not affected. The lesson is the method: one person with a swarm of agents can now test every build of a piece of software that the whole internet depends on.

A DNS filter is a suggestion until you close the ways around it.

Most unwanted traffic starts with a lookup: a tracker, a telemetry endpoint, a freshly registered domain in a phishing email. Refuse the lookup and the connection never happens.

  • I have run Pi-hole since its first release, in homes in several countries, but it only filters the devices that ask it. At StudentTenant.com's office in Romania I have run pfSense since 2015, and that is where filtering belongs: in the firewall, where every device passes and each one is seen by its own address.
  • The side doors are closed: plain DNS to other servers is redirected to the firewall's resolver, DNS over TLS is blocked, and devices with no reason to use an encrypted public resolver cannot reach one.
  • Offices and homes alike: printers, cameras, televisions, meeting-room screens and robot vacuums get their own DNS view, with only the names they need. It matters for working from home even with a VPN: a compromised device on the home network sits one hop away from the work laptop.
  • A scheduled audit once caught a vendor upgrade that had quietly removed part of a filtering stack: lookups worked, blocking did not, and nothing raised an alarm. A filter needs a test from the outside, on a schedule: look up a name that must be refused, and raise an alarm if it resolves.

Twenty servers or twenty thousand: the basics are the same

  • Count first. You cannot patch what is not on a list.
  • Decide the order before the emergency: test systems first, the critical ones last.
  • When you cannot upgrade, reduce what can reach the system, write down that it is temporary, and book the real fix.
  • Assume one will get through. Watch for the side effects: a new file, a new account, a new open port.
  • Keep outbound traffic on an allow-list, so a compromised server cannot fetch its next stage.
  • Give AI agents narrow, short-lived, logged access, with a person approving the risky steps.
  • Keep a register: every change, where it lives and how to undo it. Including the ones that went wrong.

Testing, with permission

I have run penetration tests for multiple clients over the years, including Obsentum and other companies in its group. Always under a signed scope: targets, methods, windows and stop conditions, agreed before anything starts.

fleet-watch · managed service

fleet-watch: I deploy it and run it for you.

fleet-watch is the tool I built to watch and patch the servers I am responsible for. As a managed service it does the same for yours: the alerts that matter, updates rolled out in stages, and an evidence trail, with a person behind them.

From a morning reading list to a fleet that reports on itself

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. More vulnerabilities were published in the first ten and a half months of this year than in the two years before it combined.

Then I was asked to rescue a website that had been compromised for a month. Every step of that attack had left a trace somebody could have seen at the time: a new file, a new administrator, a changed setting. A reading list tells you what could happen. Your servers tell you what did.

The first commit of fleet-watch is from the afternoon the rescued site went back online. Three weeks in October tells the whole story.

fleet-watch consolesample values
6hosts
2need you
1update job running
0pages in 24 h
  • web-01Importantup to date
  • web-02Important2 updates · 1 CVE
  • db-01Critical3 updates
  • app-01Labup to date
  • edge-fwImportantrules read-only
  • laptop-01Lab1 update

web-02

Important level · agent current · heartbeat 2 min ago

  • critical · CVEphp 8.4.26 → 8.4.27
  • securityopenssl 3.0.2-1 → 3.0.2-2
  • newer digestnginx image
Update all security

Signed job · fixed action · no shell access needed

Host names and values are made up.

How the managed service runs

  1. A read-only look

    What you run, what is exposed, what is out of date. Nothing changes.

  2. Agents and levels

    A small agent on each machine, and an agreed level for each: Lab, Normal, Important, Critical.

  3. I run it

    I watch the alerts and prepare the updates. You approve; the rollout stages itself and steps back on failure.

  4. You keep the evidence

    The event log and the change register are yours, ready for a customer or an auditor.

What it does

Finds what needs patching

It checks OS packages, containers and vendor releases against public vulnerability data, and watches new upstream builds for security fixes. Each host gets one verdict: act now, act soon, review, or fine.

Rolls the fix out

One click, then it rolls itself out: test machines first, a pause, then each level, rolled back on failure. Jobs are signed, run one fixed action and expire in 15 minutes. There is no remote shell.

Watches for a break-in

It watches for the signs of a break-in, so a break-in through a flaw nobody has published yet still shows up within seconds.

Pages only when it matters

Only a change in an attacker's favour pages a person. Everything else waits in a daily digest.

A first read by a local model

A small model on the hub reads each incident's evidence and gives a first read: likely benign, suspicious, likely malicious or unclear, and why. It never acts. You choose where the hub runs, in your network or in mine, and the evidence stays where you decide. A person decides.

Evidence you can hand over

Every event is hash-linked to the one before. One command recomputes the chain and shows the first row that does not match.

The signs it watches for

fleet-watch does not try to recognise attacks. It watches for what an attack has to do in order to succeed.

  • A new user, a new administrator, a changed permissions file.
  • A changed list of authorised SSH keys, compared by hash.
  • A new scheduled job, service or timer.
  • A new listening port, paged when it is open to the outside.
  • A new privileged binary or kernel module.
  • A new container, a privileged one, or one that can control the container engine.
  • A login from a source never seen before.
  • On web servers: a new PHP file, replaced code, mass renames, a new site administrator, a script run outside the allow-list, a login followed by file changes.

A feed can only match what already has an identifier. An exploit for an unpublished flaw has none, but it still has to leave something behind. A PHP file dropped as a test raised an alert within 15 seconds.

evidence log · hash-chainedsample values
#0411a41f0c92 ← 7be3d508 host check-in
#04125d27e8b1 ← a41f0c92 new listening port: alert
#0413e90c3a76 ← 5d27e8b1 update verified, applied

Edit any row and every later link stops matching. Verification fails.

PHP2Shell, end to end

PHP2Shell was a pair of remote-code-execution flaws in PHP. Here is how fleet-watch handles an emergency release of that class, step by step. Host names and values are sample values.

PHP2Shell, end to endsample values
[01]advisory CVE-2026-109618 and CVE-2026-109619 published; fixed in PHP 8.5.12, 8.4.27, 8.3.36, 8.2.35
[02]flag fleet-watch marks every host and container on an affected PHP build red: web-01, web-02, app-01
[03]job one signed update job to the fixed release, 8.4.27: valid 15 min, fixed action, no shell access needed
[04]lab Lab hosts first; health check after each; pause 30 min
[05]then Normal, Important, Critical in turn, one host at a time, a pause between levels
[06]failure a failed update stops the batch, pages once and puts the previous container back
[07]evidence every step written to the hash-chained evidence log

How fleet-watch handles a PHP2Shell-class release.

staged rollout · php 8.4.27sample values
Labdone
Normaldonepause 30 min before the next level
Importantrunning
Criticalqueued

On failure the batch stops and the previous version is put back.

On the hosting servers I look after, PHP was on the fixed release the morning after it shipped. The full story.

What it means for you

An agency hosting many WordPress sites

You host dozens of sites you did not write. You hear within seconds when a PHP file appears where none should be or a new administrator shows up. When PHP ships an emergency release, you see every container that needs it and roll it out test sites first.

A SaaS company with a small team

Nobody on your team has time to read advisories. Each server gets one verdict, updates go out in stages with a step back if a service fails, and you are paged only when something changed in an attacker's favour.

A regulated supplier

Your customer asks how you know a server was not changed, and how fast you patch. Every event is hash-linked and verifiable, every update is a signed job with a record, and you choose where the evidence lives: the hub runs in your network or in mine.

An online shop before its peak season

The patches nobody dares apply before a sales peak go to a test machine first, wait, and roll back on their own if something fails. You go into the peak patched, not hoping.

Running today, and what is next

Running today

On Linux servers and containers: break-in detection, the vulnerability and update view, signed update jobs, staged rollout with rollback, the local first read and the hash-chained log, with the build-watcher reading each new upstream build for security fixes. Update jobs are switched on host by host.

Next

Windows and firewall agents are built and next to run on real machines. The firewall agent only observes; it never edits a rule.

Designed

One hub for several clients, each with a separate encrypted store. Hardening packs that are generated and checked, never applied on their own.

Need patching that keeps up?

Send a short message: what you run, what worries you, any deadline.

Get in touch