peekgenticBook a call

News analysis · AI security

Google: AI now turns security patches into attacks with no human. What it means for your office

Google's threat team says it found an AI system that reads software patches, writes the attack, and runs it on its own. The gap between a patch and an attack is shrinking. What changes for a 50-300 person company, what doesn't, and what to check this week.

By Miguel GutierrezCo-founder, Peekgentic · Dallas
· 4 min read
Preview card: AI now turns patches into attacks with no human.
The attack methods are old. The speed is new.Peekgentic

In brief

On Nov. 16, Google's Threat Intelligence Group reported finding, for the first time, attackers using AI to fully automate finding and exploiting security holes with no human in the loop. The system compares software before and after a security patch, works out what was fixed, writes an attack, and runs it against websites that haven't patched yet. For your office, the lesson isn't new, but the clock is faster: anything you have on the internet (website plugins, firewalls, VPNs, print servers, portals) needs patches in days, not at the next quarterly maintenance window.

  • Google says attackers now use AI to turn security patches into working attacks with no human in the loop.
  • In the PaperCut case, an AI-run attack began four days after the advisory and reached network admin in about six hours.
  • Anything your company has on the internet needs patches in days, not on a quarterly schedule.
  • Old gaps still matter: two unpatched 2021 Windows flaws helped the PaperCut attacker take over networks.
  • Basic hardening still works: a web application firewall stopped one attack.

What Google found

Google's Threat Intelligence Group and Mandiant publish a regular report on how attackers use AI. The Nov. 16 edition says that in the past quarter, for the first time, it saw attackers use AI to fully automate finding and exploiting software flaws, without a human in the loop.

The details are specific. Google found an exposed server running a system it calls exploit_agent, wired to several AI models. It watches popular web software, compares the code before and after a patch is announced, figures out what was fixed, drafts an attack, and runs it. Google says the operator mostly went after French media, entertainment and education websites. One example in the report is a WordPress plugin.

The report's broader point: attackers are using AI to "compress attack timelines," shrinking the window between a patch being released and an attack that uses it.

What that looks like in a real attack

An earlier case shows the speed. In late August, print management software PaperCut disclosed two serious flaws. Security firm GreyNoise tracked what happened next: four days after PaperCut's advisory, an attacker used AI agents to download the old and new versions, compare them, build a working attack, test it in a lab and launch it.

The PaperCut campaign, in hours

GreyNoise's timeline of an AI-run attack. Shorter is worse for defenders.

  1. Vendor advisory to attack work starting4 days (96 h)
  2. Empty workspace to first break-in at a real victimjust under 4 h
  3. First break-in to first network admin rightsabout 2 h more
  4. Fastest takeover at one victim, access to network admin7 minutes

Source: GreyNoise, Sept. 9, 2026. Once the campaign launched, it reached 11 organizations in 26 seconds.

In all, GreyNoise counted at least 440 PaperCut servers at 395 organizations in 48 countries. The attacker reached full network admin rights at 12 of them. At one US high school, it went from first access to full control in seven minutes.

Notice what PaperCut is: a print server. GreyNoise notes that its self-hosted versions run with top-level Windows rights by default and are usually tied into the company's Active Directory. That's the kind of quiet box in the back room that nobody thinks of as "internet software," until it is.

The attack isn't new. The speed is.

What changes for a mid-size company

Most 50-300 person companies patch on a schedule: a monthly Windows cycle, website updates when someone gets to them, firewall updates once a year. That rhythm assumed attackers needed weeks to turn a patch into a working attack. Both reports say that's no longer safe for anything facing the internet.

For a benchmark, the federal government's own rule is a useful yardstick. CISA's directive BOD 22-01 requires federal agencies to fix flaws on its known-exploited list within two weeks for newer flaws. You aren't bound by it, but if a federal agency gets two weeks, a company with a public-facing print server shouldn't take two months.

The old stuff still matters too. GreyNoise found that at some victims, the attacker got from the print server to full control by using two Windows Active Directory flaws from 2021 that had never been patched. AI made the first step faster. Old gaps did the rest.

What doesn't change

Here's the fair side. Google's report says these AI attack tools are fast but noisy: they generate traffic and leave traces that make them easier to spot. Google also found that hundreds of AI-written attack scripts posted after a recent PHP patch didn't actually work. And GreyNoise saw one attack fail against a site behind Cloudflare's web application firewall. Its takeaway: basic hardening still works against AI-run attacks.

Most of the attacks also used old methods: stolen passwords and unpatched systems. Google says the basic mechanics of these break-ins haven't changed. That's good news. It means you don't need a new kind of defense. You need the existing basics done faster and more completely.

What to check this week

Five questions for your IT person or provider

AskWhy it matters now
What do we have that's reachable from the internet?Website and its plugins, firewall, VPN, remote access, print server, customer portal. You can't patch what's not on the list.
How fast do internet-facing systems get patched?Aim for days. Turn on automatic updates for website plugins where you can.
Does anything internal face the internet that shouldn't?Print, file and admin servers usually don't need to be public. Put them behind the VPN.
Is there a web application firewall in front of our site and portals?One stopped an attack in the PaperCut case.
Are there old Windows patches still missing on servers?Two 2021 flaws let the PaperCut attacker take over whole networks.

If you use a managed IT provider, ask for the answers in writing, with the date each internet-facing system was last updated. If your website runs on WordPress, ask who updates the plugins and how often. Plugins are exactly the kind of popular web software Google says the AI system was watching.

For more on how AI is changing the patch side of this, see our earlier pieces on AI-found flaws and patching and the PHP flaws AI found.

Common questions

What did Google's threat team report on Nov. 16, 2026?

Google's Threat Intelligence Group said it saw attackers use AI to fully automate finding and exploiting software flaws, without a human in the loop, for the first time. A system it calls exploit_agent compares software before and after patches, writes attacks and runs them.

How fast can attackers use a new patch now?

In the PaperCut case tracked by GreyNoise, an AI-run attack started four days after the vendor's advisory. It went from an empty workspace to a first break-in in just under four hours, and to network admin rights about two hours later.

What should a mid-size company patch first?

Anything reachable from the internet: website plugins, firewalls, VPNs, remote access, print servers and customer portals. Aim to patch those in days, and take internal servers off the public internet.

Does basic security still work against AI attacks?

Yes. GreyNoise saw a web application firewall stop one attack, and Google says the attacks still rely on stolen passwords and unpatched systems. AI makes attacks faster, not different.

What is the federal deadline for patching known exploited flaws?

CISA's directive BOD 22-01 requires federal agencies to fix flaws on its known-exploited list within two weeks for newer flaws. Private companies aren't bound by it, but it's a useful benchmark.

Are AI-written attacks always effective?

No. Google found hundreds of AI-generated attack scripts after a recent PHP patch that didn't work, and says current AI attack tools are noisy and leave traces.

Sources

  1. Google Threat Intelligence Group, Agentic Threat Activity Crosses a Significant Threshold (Nov. 16, 2026)
  2. GreyNoise, Agents Gone Wild: An AI-Orchestrated Global Campaign Against PaperCut NG/MF (Sept. 9, 2026)
  3. CISA, BOD 22-01: Reducing the Significant Risk of Known Exploited Vulnerabilities

More from Peekgentic

Published November 17, 2026 · Written by Miguel Gutierrez, Co-Founder, Peekgentic (Dallas, TX). Every statistic links its source.