AI and security
Google says attackers now use AI to turn software fixes into attacks with no human involved
Published November 19, 2026
Google’s security researchers say that, for the first time, they caught attackers using AI to run the whole process of breaking into websites without a person stepping in: reading a software fix, working out the flaw it repairs, writing an attack and launching it. The practical lesson is an old one with new urgency: the time between a fix being published and criminals using it is shrinking, so updates matter more than ever. These are Google’s own findings.
What happened
On November 16, the Google Threat Intelligence Group (GTIG) and its Mandiant incident-response team published their quarterly report on how attackers use AI. The headline finding: in the past quarter, GTIG says it identified attackers using large language models (LLMs, the kind of AI behind chatbots) to find and exploit security flaws with no human in the loop.
The evidence was an exposed “staging server,” a computer the attackers used to prepare their work, that anyone could reach. On it sat a system the
attackers called exploit_agent. It is a multi-agent framework, meaning several AI programs that each handle one step and pass work to each other.
Google says it worked with several AI models and kept comparing popular web software before and after each security update. When a company ships a patch
(a fix for a security hole), the change itself shows where the hole was. The system read those changes, guessed at the flaw, drafted an exploit (code that
takes advantage of the flaw) and ran it automatically.
Google also found scripts for what attackers do after getting in, such as looking around inside a network, guessing passwords and planting a webshell (a hidden back door on a web server). Data on the server suggested the operator went after French websites in media, entertainment and education.
TechTarget notes that earlier editions of Google’s tracker had only seen partly automated attacks.
Why it matters to a normal person
Security people split flaws into two kinds. A zero-day is a hole nobody has fixed yet. An n-day is a hole that already has a fix, but plenty of computers have not installed it. Google says attackers may find it more efficient to use AI on n-days: they only have to study a small, known change instead of searching a whole program for something new.
Google’s numbers show the pressure. It says overall flaw disclosures and exploits nearly doubled between January and August 2026. The report also points to an August attack on PaperCut, office printing software, where security firm GreyNoise reported a campaign against hundreds of organizations four days after the flaws were disclosed, likely with AI-built exploits.
Our advice: turn on automatic updates for your phone, computer, browser and home router, and install security updates soon after they arrive rather than next month. If you run a website on WordPress or a similar system, keep its add-ons updated too. Google’s report shows the attack system analyzing a popular WordPress plugin. Jason Elrod, security chief at MultiCare Health System, told TechTarget that some fixes will need to happen “in hours rather than days or weeks.”
What it does not mean
It does not mean AI has broken computer security. Google says the methods are the same as before, such as stolen passwords and unpatched systems. What changed is speed and scale. Last month we covered two PHP flaws that AI helped find, and Google says hundreds of fake, non-working attack demos, likely AI-made, appeared after those were patched. Plenty of AI-written attacks simply fail.
It also does not mean these tools are invisible. Google says today’s automated attack tools “trade stealth for velocity”: they create a lot of traffic and leave clear traces, which makes them easier to catch. Google expects that to improve as the AI improves.
Finally, keep the source in mind. This is one company’s view, built on what its own systems and customers see, and Google sells security services. The report describes one exposed server and does not say how many sites that system broke into.
Sources
- GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, Google Threat Intelligence Group (Nov 16, 2026; primary)
- Google finds AI agents autonomously turning patches into exploits, Alissa Irei, TechTarget (Nov 17, 2026)