Talk to us

Blog · 17 November 2026 · Rob Sherwood

AI Can Now Turn a Security Patch Into an Attack

Two bits of security news landed within a few days of each other this week, and together they change how I'd think about updates on any bespoke app. One is about attackers. The other is about the people writing the fixes. Both involve AI doing work that used to need a skilled person and a fair bit of time.

The short version: the gap between "a fix is published" and "someone is using that bug against you" is getting smaller, and it's being made smaller by software rather than people.

What Google found

On 16 November, Google Threat Intelligence Group published its latest AI threat tracker. The headline finding is that it saw attackers using AI frameworks to automate finding and exploiting vulnerabilities, in Google's words, "without a human in the loop for the first time".

The tool it describes is called exploit_agent. Google calls it a multi-agent framework that plugs into several AI models, both big commercial ones and models run locally. What it does is simple to describe. It compares the code before and after a security patch is announced, works out what the patch fixed, writes an exploit for it and then runs that exploit against real targets. Automatically. Google says data was taken from victims, and the scripts it found included ones for installing webshells (a small backdoor left on a server so the attacker can come back).

That first step deserves a moment. When an open source project fixes a security bug, the fix is public. It has to be, because everyone needs to install it. Working backwards from a fix to the bug it closed has always been possible. It just used to take an experienced person hours or days. If a machine can do it unattended, the head start you used to get between the fix going out and attacks starting is shrinking.

What Canonical is doing about it

On 12 November, Canonical (the company behind Ubuntu) announced Surge, which it describes as "a multi-agent orchestration framework for CVE remediation and backporting". In plain English, AI models that take known security fixes and work them into the older package versions that each Ubuntu release ships.

The interesting part is how careful they've been. Several competing models each produce a fix. A separate panel of judge models has to agree. Anything low confidence, unusual or deadlocked goes to Canonical's human security engineers, and everything goes through build and deterministic tests before it's published. Even then, Surge fixes only go into a new opt-in "Edge archive" for Ubuntu Pro users. Canonical is clear that it will not automatically upload these fixes to its stable releases, and that normal Ubuntu security updates "remain strictly human-validated". The kernel isn't part of it either.

Canonical also says it is shipping double the number of security fixes it did last year. That's the other half of the story. There are simply more fixes to get out, and more of them being found with AI's help.

Should you switch on the Edge archive?

For a production server running a client's business system, I wouldn't. Canonical itself calls it "an opt-in preview of fixes in development". That's a perfectly sensible thing to try on a test box, and Ubuntu Pro is free for personal use if you want to have a look. But the whole point of a stable release is that changes have been through proper testing, and the normal security channel is still getting the fixes out.

The bit I'd take away isn't Surge itself. It's the direction. Attackers are using AI to read fixes faster, and defenders are using AI to write them faster. Whoever applies them slowest is the one left exposed.

What this means for a bespoke app

Your operating system is only part of the picture. A typical custom web app also sits on a language runtime (PHP, Node, Python, .NET), a framework, a web server, a database and a long list of open source packages pulled in through Composer, npm or pip. Every one of those publishes its security fixes in public. Every one of them is exactly the kind of public fix a tool like exploit_agent can read.

So a few practical things worth checking:

  • Do you know what's in it? Someone should be able to list the runtime version, the framework version and the main packages, and say which are still supported. If nobody can, that's the first job.
  • Are OS security updates applied automatically? On Ubuntu, unattended security updates are there for a reason. They should be on, with someone checking that reboots for kernel fixes actually happen.
  • Who watches the app's own dependencies? The OS updater won't touch your Composer or npm packages. Somebody has to run the audit tools, read the advisories and ship the updates, and that means a developer who knows the code well enough to know what an update might break.
  • How long does a fix take to get live? If deploying a one-line package update means a week of nervous manual testing, that's now a security problem as well as a productivity one. Decent tests and a repeatable deploy are what let you move quickly.
  • Is it stuck on something old? An app pinned to an unsupported framework version can't take the fix at all, however fast you are. That's where a planned upgrade earns its keep.

None of this is new advice. What's changed is how much it costs to be a few weeks behind. I wrote on my own site about patching around 100 sites overnight for PHP2Shell with AI helping, and the lesson there was the same one: the tools make you faster, but someone still has to know the systems and be awake to do it.

Where we come in

If you've got a bespoke app and you're not sure who's keeping it up to date, our code audit is a sensible starting point. It's £250 + VAT, covers the code, the server and the database, and tells you plainly what's out of date and what to fix first. If the server is the worry, my server health check looks at updates, access and backups in the same plain way.

If you'd rather someone just took care of it, that's what Look after it is for: security updates for frameworks and libraries, hosting, tested backups and a person who picks up the phone. It starts from £355 a month + VAT. We're also happy to take on existing software that someone else built, including apps built with AI.

Either way, drop us a line and we'll have a straight conversation about where your app stands.

Cheers, Rob

Tell us what you've got

It goes straight to Rob and Jase. You'll hear back from one of us, usually the same day, and we'll tell you honestly whether we're the right people.

Just want it checked? Code audit, £250 + VAT