Construction AI BriefSubscribe →
Issue
№447
Pillar
Trend
Audience
Trade sub
Dated
2026.11.17

Google found an AI tool that turns security patches into attacks on its own. A sub should ask its IT provider how fast patches go in

Google's threat team found an attacker framework that reads each new software patch, works out the flaw it fixed, and attacks sites that haven't updated, with no person involved. For a trade sub with a WordPress site, a print server and remote access to the accounting box, the question for the IT provider is how many days a patch waits.

ByConstruction AI BriefAbout this publication

Attackers now have software that reads a security patch, figures out what the patch fixed, and attacks everyone who hasn't installed it yet, with nobody at the keyboard. For a trade sub, that makes "how many days do our patches wait?" a question worth putting to the IT provider this week.

The finding comes from Google's Threat Intelligence Group, in a report published November 16.

What did Google find?

On an exposed staging server, Google found a framework its operator called exploit_agent. Google describes it as connected to "multiple frontier and local LLMs" and set up to do four things in a loop:

  1. Compare the code of popular web applications before and after each patch announcement
  2. Analyze the difference to identify the vulnerability
  3. Draft an exploit
  4. Run it "without human intervention"

The code sample Google published shows the framework examining access controls in Spectra, a WordPress plugin. VentureBeat reports Google calls this the first time it has seen attackers fully automate discovery and exploitation with no human in the loop.

Reading a patch to find the hole is an old technique. As Merritt Baer, a former deputy CISO at AWS, told VentureBeat: "Attackers didn't invent a new kind of bug. They automated the tedious part."

How fast is fast?

Google did not measure a patch-to-exploit time for exploit_agent. The report's hard numbers come from a separate case involving PaperCut, software that manages office printers:

  • Researchers and the security firm Huntress built working exploits for two chained PaperCut flaws within a day of disclosure
  • A campaign that Google says likely used AI models to build and test its exploits started four days after disclosure and hit hundreds of organizations
  • In that campaign, per GreyNoise, the attacker went from a first foothold to domain administrator "within minutes"

Google's conclusion is that this kind of tooling is shrinking the patch window. It also suggests attackers may find it more efficient to aim AI models at already-patched flaws than to hunt for unknown ones, because a patch hands them a short, finite set of code to study.

What does a trade sub actually have exposed?

More than most owners think, and almost none of it is watched daily. A mechanical or electrical shop with an office and a few remote users typically has some mix of:

SystemWho usually patches itWhat to ask
Company website (often WordPress with a dozen plugins)A marketing freelancer, or nobodyAre plugin auto-updates on? Who gets the alert when one fails?
Firewall and VPN for remote accessThe IT providerHow many days after release does firmware go in?
On-prem accounting or estimating server reached from outsideThe IT providerIs it reachable from the internet at all, or only through the VPN?
Print management and plotter softwareWhoever installed itIs it exposed to the internet? It should not be.

The website is the one that gets forgotten. It was built once, it works, and its plugins sit several versions back. That is the kind of software the Google sample shows the framework studying.

A "patch on the second Tuesday, after testing" routine was built for a world where turning a patch into an attack took a skilled person days or weeks. The PaperCut case shows one day for a working exploit and four for a mass campaign.

What is still unproven?

Quite a bit, and Google is upfront about it:

  • No specific vulnerability is tied to a successful exploit_agent attack
  • No victim count is given
  • Google could not show the framework caused the compromises whose data sat on the same server
  • The operator's known interest was French media, entertainment and education sites, not contractors

So this is not evidence that anyone is aiming such a tool at construction firms. But a tool that attacks every unpatched copy of a plugin it can find does not check what trade the site owner is in. And faster patching does not help a system that was entered before the patch went on, which is why Baer's advice to VentureBeat included monitoring for active exploitation, not just missing patches. We covered a related exposure in homemade web apps yesterday.

Takeaway: Send your IT provider three questions. What do we run that is reachable from the internet? How many days does a critical patch for each item wait, on average? Who looks for signs of a break-in after we patch? If the answer on the website is "not ours," turn on automatic plugin updates today and delete the plugins nobody uses. Anything that does not need to face the internet, starting with the print server and the accounting box, goes behind the VPN.

Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com and forward this to whoever signs your IT provider's invoice.

FAQCommon questions
What is exploit_agent?
It is an attacker-built framework that Google's Threat Intelligence Group found on an exposed staging server and described in a report published November 16, 2026. Google says it connects to several AI models, compares popular web application code before and after each patch announcement, identifies the likely flaw, drafts an exploit and runs it without human intervention.
Did exploit_agent actually break into anyone?
Google has not confirmed that. The server also held victim data and post-compromise scripts, but Google did not name a vulnerability the framework exploited, give a victim count or report a measured patch-to-exploit time. A comment in its configuration file describes it as a pilot.
How fast are attackers using AI to exploit newly disclosed flaws?
In the one case Google's report puts numbers on, two flaws in PaperCut print management software, researchers and the security firm Huntress generated working exploits within a day of disclosure. A campaign that Google says likely used AI models to build its exploits began four days after disclosure and hit hundreds of organizations.
Does this affect a construction company's website?
It can. The code sample Google published shows the framework inspecting a WordPress plugin, and Google says the operator's known interest was French media, entertainment and education sites. Automated attacks on a known plugin flaw do not sort targets by industry, so a contractor site running an outdated plugin is exposed in the same way.
What should a subcontractor ask its IT provider about patching?
Ask for a list of everything the company runs that is reachable from the internet, how many days a security patch for each one normally waits before it is installed, and whether WordPress plugins update automatically. Then ask who watches for signs of a break-in, since a patched system can already have been entered.
End of sheet — issue №447
Published · 2026.11.17
Project
Construction AI Brief
Dated
2026.11.17
Sheet
1 / 1
Rev
A
Published independently · constructionaibrief.com · © 2026Facebook·Privacy·About