Google Just Caught an AI Agent Turning Security Patches Into Working Exploits — and the Example Is a WordPress Plugin on 1 Million Sites

Google's threat team just published something I have never seen before: an AI agent framework that reads a security patch, works out what was fixed, writes an exploit for it, and fires it — with nobody at the keyboard. The example in Google's own screenshot is a WordPress plugin running on more than a million sites. Here is what actually happened, what it does not prove, and the checklist I am running on my own sites this week.

Google Just Caught an AI Agent Turning Security Patches Into Working Exploits — and the Example Is a WordPress Plugin on 1 Million Sites
Google Just Caught an AI Agent Turning Security Patches Into Working Exploits — and the Example Is a WordPress Plugin on 1 Million Sites — Technical Architecture & Overview

Table of Contents

I have built WordPress sites with Elementor for years, and most security advisories follow the same shape: a flaw is found, a patch ships, you update, you move on. The one Google published on November 16, 2026 broke that shape completely.

Google's Threat Intelligence Group released its quarterly AI Threat Tracker, "Agentic Threat Activity Crosses a Significant Threshold." Its headline finding is a first: last quarter, Google identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation — without a human in the loop. An agent that does the whole chain itself. And the example Google chose to print is a WordPress plugin installed on more than a million sites.

What Google actually found

Google says it identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework the company calls exploit_agent, integrating multiple frontier and local large language models. The same server held reconnaissance, brute-forcing and webshell scripts, plus victim data. A configuration file carried Russian-language comments, one translated by Google as "LLM API keys for auto-exploitation (exploit_agent pilot)."

The framework runs what Google calls continuous differential analysis. In plain English: a vendor publishes a fix; the agent compares the software before and after the patch; it works out what vulnerability the patch closed; it drafts an exploit for that hole; and it automatically executes it against sites still running the old version. Google's words: the framework drafts an exploit and "automatically executes them without human intervention." The patch itself is the only input, and every patch comes with one.

Google is careful about its limits, and so am I. The report gives no success rate, no count of sites got into, and no name or country for the operator. The Register says only that the operator is suspected Russian-speaking; victim telemetry points to French media, entertainment and education websites. The mechanism is established. The body count is not.

Why the example is a WordPress plugin

Figure 4 of Google's report is captioned: "Agentic framework inspecting access control implementations in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25." The WordPress.org listing for that plugin, now Spectra Legacy, shows 1M+ active installations, version 2.20.6 dated November 3, and a changelog in which the last three releases — September 28, October 9, November 3 — each fixed a security issue reported by Wordfence or Patchstack.

2.19.25 is the last version before 2.19.26, which in May closed a flaw rated 8.8 out of 10 that let a low-privilege Contributor user run their own code on the server, according to ZeroPath (CVE-2026-7465). Google does not say whether the agent was looking for that flaw or another, or found nothing at all.

It makes sense once you know the difference between a zero-day and an n-day. A zero-day is a flaw nobody knows about; finding one means casting a wide net and hoping. An n-day is a flaw that has already been fixed: the fix is public, the old code is public, and the difference between them describes exactly where the hole was. Google's argument is that threat actors may find it more efficient to use LLMs to rapidly weaponize patched n-days, because they can focus on a finite set of source materials. A model does not need to be brilliant to read that diff — only patient, in a way no human team can be. WordPress is the biggest feedstock of all.

How fast the window has become

You used to have days or weeks after a patch shipped. That model is already dead. The report's clearest timeline is from late August: for two flaws in PaperCut, working exploits appeared within a day of disclosure, and a scanning campaign against hundreds of organisations began four days later. Google says the attackers went from first foothold to domain administrator within minutes. That was with humans still steering — exploit_agent removes the day.

Disclosures and exploits nearly doubled between January and August 2026, and zero-days, which averaged eight a month in 2025, averaged eleven a month this year and spiked to twenty-two in August. BleepingComputer, reading the report's timeline graphic, puts Google's estimate at mostly autonomous unsophisticated attacks within three months and sophisticated fully autonomous ones within six — a trade outlet's reading, not a Google quotation, but the order of magnitude is months, not years.

When the PHP project fixed two flaws on October 21, hundreds of claimed LLM-generated proof-of-concept exploits appeared, most of them broken — and a site owner cannot tell a working exploit from a broken one by looking. One trade-off favours you: these frameworks trade stealth for velocity. An agent that tries an exploit the day a patch ships is noisy — it scans widely, fails often, and leaves logs. A patched site sees nothing but noise; an unpatched one is what the noise was for.

Three days earlier, Amazon reported the same pattern: its decoys caught ARTEX agents stealing provider API keys from exposed AI tooling — LiteLLM, One API, Dify, Langflow, Ollama — mostly through default keys never changed and admin pages without passwords. Its advice: watch for unexpected spend, because that is what a stolen key looks like from the inside. None of these techniques were new. What agents change is the speed and persistence of the attacker: the patch is now the recipe.

What this means if you build with Elementor

Sort what you run into two piles. The first is software somebody else patches the moment a fix exists — a hosted checkout, a managed CDN, a hosted store's core platform. Nothing here changes your position there, because that window is not yours to close. The second pile is everything you installed yourself and must update: a WooCommerce store and its plugins, the WordPress blog beside a storefront, the landing-page site the agency built two years ago, and the AI tooling someone set up to try an idea. The moment a vendor in the second pile ships a fix, the fix becomes the map, and the race to apply it runs against something that does not sleep.

Two common mistakes, both now wrong. First, "we are on a hosted platform, so this is the platform's problem." Your storefront may be; the rest is not. The agent does not care which of your properties takes the orders — only which one runs last month's version of a popular component. The report's victims were media, entertainment and education websites: content sites, not checkouts. Second, "we will read the advisory and decide whether it applies." That worked when the attacker needed a week to turn an advisory into an attack. Today's report removes the week: the agent reads the code, and it has started before the advisory is written. A three-business-day review is now a three-day window with your name on it.

For the second pile, decide once, in advance: updates apply automatically, review happens afterwards on a staging copy — except where an update could break checkout, where staging updates first and a person checks it the same day. The encouraging part: the agent and you are reading the same document. If your site is on the fixed version by the time the agent's draft is ready, the agent has nothing to run. Automatic updates, a staging copy, and a short list of what you actually run are the finish line, and it is close.

What to do this week

  1. Write down the second pile. Every site, plugin, theme and self-hosted tool your business runs that somebody on your side has to update — including the ones the agency built. If the list has to be assembled from memory, that is the finding.
  2. Turn on automatic updates for plugins and themes on every WordPress property. WordPress has offered per-plugin auto-updates since version 5.5. Pair it with a daily backup you have actually restored once.
  3. Check the plugin in the report, and the ones like it. If any of your sites runs Spectra, confirm it is on 2.20.6 or later, then look for anything with a security fix in its changelog in the last ninety days — those patches have already been read by someone.
  4. Give the side sites a staging copy. A clone that takes updates a day before production, checked the same day, turns automatic updates from a risk into a routine — and for WooCommerce it is the only safe way to let payment and shipping plugins update themselves.
  5. Find the AI tooling. Search your cloud bill and DNS for Amazon's list — LiteLLM, One API, Dify, Langflow, RAGFlow, Open WebUI, Ollama and relatives. For each, confirm it is behind a login, the default admin key is changed, it is not internet-reachable unless it must be, and the key inside has a spending limit and an alert.
  6. Set spend alerts on every AI provider account. A stolen key gets used, and the first sign is a bill. An alert at twice last month's usage costs nothing and arrives in hours, not at month end.
  7. Look at your access logs for the week after any plugin fix. A burst of requests to a plugin's endpoints from unfamiliar addresses in the days after a fix ships is what an automated attempt looks like — visible in ordinary logs.

None of these steps is difficult. The difficulty is that nobody owns them: the store team owns the storefront, the agency owned the blog until the contract ended, the developer who set up the AI gateway has moved on, and the host patches the server but not what runs on it. The defence needs one person holding the inventory, setting the update policy, confirming staging and alerts, and reviewing the list monthly.

Conclusion

WordPress updates used to be good hygiene you did eventually. Google's November 16 report changes the tense. The attacker is no longer a person who might get around to your site. It is a loop that starts the moment a patch is published, reads the diff, writes the exploit, and fires it at every site still on the old version, without pausing. The defence is the same, only faster: know what you run, update the day the fix ships, keep a staging copy and a backup. The patch is the attacker's recipe now — make sure it is also your shield, applied before the agent's draft is ready. If you run WordPress with Elementor, start with step two today — ten minutes.

Frequently Asked Questions

What is exploit_agent?

A multi-agent framework Google found on an exposed staging server. It compares software before and after a patch, drafts an exploit for the hole the patch closed, and executes it automatically against sites still on the old version.

Did Google prove that exploit_agent hacked the Spectra plugin?

No. Google's report shows the framework inspecting access-control code in Spectra 2.19.25, and describes what the framework is built to do with what it reads. Google does not say the agent found a flaw in Spectra, exploited it, or compromised any site — no success rate, no count of sites got into.

What is an n-day vulnerability, and why does it matter now?

An n-day is a flaw that has already been patched. The fix is public, the old code is public, and the difference between them describes exactly where the hole was. Agents only need to read that diff, not hunt for unknown zero-days — so the window between a patch shipping and your site updating is now the whole game.

I use Elementor — does this affect me?

Elementor itself is not named in the report, but Elementor sites run on WordPress, and WordPress sites run plugins. The plugin in Google's screenshot, Spectra, has more than a million installations. If your site runs Spectra, or any popular plugin with a recent security fix, update promptly — or the patch becomes a map to your site.

Sources: Google Threat Intelligence Group (16 Nov 2026); The Register; BleepingComputer; Amazon Threat Intelligence (13 Nov 2026); WordPress.org; ZeroPath.

Looking for more WordPress & Elementor tutorials?

Explore our complete directory of 53+ web performance, theme development, and Elementor architecture guides.

View Topic Sitemap XML Feed