Guide · AI security · as of 18 November 2026

AI agents now turn security patches into attacks. Five checks.

When a software maker fixes a security hole, it publishes the fix. The fix also shows where the hole was. Google’s threat intelligence team reported this week that it found an attack system doing that reading on its own: it compares each new version of popular web software with the old one, works out the hole, writes an attack and runs it, with no person involved. Google calls this the first case it has seen of attacks found and carried out fully by AI without a human in the loop. For any business with a website or web app, it means the days between “fix released” and “fix installed” are now the dangerous days. This guide covers what Google found, what we found when we measured our own delay, and five checks.

What Google found

A server that reads every fix.

Google’s team found a staging server left open on the internet. It held a system its operator called exploit_agent, built from several AI agents working together and wired to both commercial and self-hosted AI models. It was set up to watch popular web software and compare the code before and after each security fix. From the difference it works out the weakness, then writes and launches an attack. The server also held data taken from victims, and scripts for what attackers do once inside, such as guessing passwords and planting hidden back doors. Google says the victims the operator was interested in were French media and entertainment websites and the education sector. One screenshot shows the system studying the access rules in a popular WordPress add-on, the kind of plugin thousands of small business sites run. Google describes the operator as suspected Russian-speaking and does not name any victim.

Why the window is shrinking

A fix is also a map.

A hole that has been fixed but not yet installed everywhere is called an n-day. Finding a brand new hole means searching everything. Attacking an n-day means reading one small change, and that is work AI does well. Google gives two recent examples. In late August, researchers used AI to build working attacks on two PaperCut print server holes within a day of their disclosure, and four days after disclosure a criminal campaign, likely built the same way, hit hundreds of organisations. In October, AI-assisted research found two PHP holes, and once the fixes shipped on 21 October, hundreds of fake attack demos, likely AI-written, appeared online. Google also counts the pace. Disclosures and exploits nearly doubled between January and August 2026. Attacks on holes nobody had fixed yet went from an average of 8 a month in 2025 to 11 a month this year, and 22 in August alone. The attacker’s side is getting faster. The defender’s side is still a person who has to notice the fix, test it and ship it.

The FBI’s alert on an AI attack agent that looks for forgotten apps

What we found in our own apps

Fifteen days, and a template.

On 3 November 2026 Next.js, the framework most of our sites are built on, published fixes for seven holes, one rated high. On 18 November we checked the Next.js version in every one of our 32 Next.js code projects. Five were on the affected version line: four live sites and apps, and the starter template we copy every new site from. We moved all five to the fixed version the same day and rebuilt each one. From fix released to fix installed: 15 days. Two lessons we did not expect. The first was the template: every site started from it would have shipped with the old version and nobody would have looked. And the high-rated hole needs one specific caching feature; one of the five uses it, only for public blog posts, so what could have leaked was already public. That was luck of design, not speed. We did not search our logs for signs of an attack, so we do not claim there were none.

Five checks for your websites and web apps

Run these this week.

  • 1. Keep a list of every site and app you run, with the software and version under it. Include old campaign sites and the template your agency copies from. Test: pick one framework or plugin and answer, from the list alone, which of your sites run it and on which version.
  • 2. Set a clock for fixes and measure it. For example: internet-facing and rated critical or high, installed within two working days; the rest within two weeks. Test: take the last security fix for your main site and count the days from its release to it being live.
  • 3. Make sure someone hears about each fix the day it ships. Subscribe to the security notices of each framework and plugin you use, or turn on automatic alerts in your code host. Test: name the person who receives them, and the last notice they acted on.
  • 4. Trust only the vendor’s word on what a hole does. After the PHP fixes, hundreds of fake attack demos circulated. Read the vendor’s notice for what is affected and what fixes it, and judge your exposure from that. Test: for your last fix, find the vendor page that says which versions were affected.
  • 5. Remove what you do not use. Every unused plugin, old test site and spare admin page is something an attack system can read and target. Test: list the plugins or add-ons on your main site and delete any nobody can name a reason for.

Questions

Questions about AI and patching.

  • What is exploit_agent?

    The name of an attack system Google’s threat intelligence team found on an exposed server and described on 17 November 2026. Several AI agents compare popular web software before and after each security fix, find the weakness, write an attack and run it without a person involved.

  • What is an n-day vulnerability?

    A security hole that the software maker has already fixed, but that is still open on every system where the fix has not been installed. A zero-day is a hole with no fix yet.

  • What is patch diffing?

    Comparing the code of a program before and after a security fix to find exactly what changed. It points straight at the hole. Defenders use it to understand a fix; attackers use it to build an attack for systems that have not installed it.

  • How fast should a business install security patches?

    There is no single legal rule for private businesses. A practical target is two working days for internet-facing systems where the fix is rated critical or high, and two weeks for the rest. Whatever target you pick, count the days on your last fix so you know where you stand.

  • Are WordPress sites at risk?

    Google’s report shows the attack system studying a popular WordPress add-on. Sites with many plugins have many fixes to keep up with. Keep plugins updated, remove unused ones and turn on automatic security updates where you can.

  • Does this matter to a business in the UAE?

    Google describes the targets as French websites and names no victim, in the UAE or anywhere else. The method reads public fixes for software used everywhere, so any business running a website or web app on common software faces the same race.

Sources

read 18 November 2026

Check it at the source.

  1. Google Threat Intelligence Group: AI Threat Tracker, Agentic Threat Activity Crosses a Significant Threshold (the report of 17 November 2026 with exploit_agent, the PaperCut and PHP examples and the monthly counts)
  2. Next.js: November 2026 Security Release (the fixes of 3 November 2026 we measured our own delay against)
  3. BRev Digital: AI agents are stealing AI keys to pay for their own attacks (Amazon’s report on attack agents from the same month)

A question?

How long do your fixes take?

Tell us which sites and apps you run and who builds them. We will help you list what sits under each one, measure your last fix, and say plainly what to update first.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com