Google Finds First Fully Autonomous AI Attack System Turning Security Patches Into Exploits

Google's threat intelligence team says attackers have, for the first time, used an AI framework to turn security patches into working exploits with no human in the loop. The system, called exploit_agent, watches popular web software for fixes, works out what was patched, and fires the exploit itself — shrinking the time defenders have to update.

Nov 18, 2026 - 02:41
 0  3
Google Finds First Fully Autonomous AI Attack System Turning Security Patches Into Exploits
A security analyst in a suit works on a laptop at a desk, with computer monitors showing lines of green code behind him.

Google Finds First Fully Autonomous AI Attack System Turning Security Patches Into Exploits

Google's threat intelligence team says it has now seen something it had been watching for: attackers using an artificial intelligence framework to find software flaws and exploit them with no human involved at any step. The finding, published on 16 November 2026, marks the first time Google has observed a fully autonomous attack pipeline operating against real targets.

The system is called exploit_agent. According to the Google Threat Intelligence Group (GTIG), it was discovered on a server the operator had left exposed to the internet, wired into several frontier and locally hosted large language models. Its job is to keep comparing the code of popular web applications before and after each security patch, work out what the patch fixed, draft an exploit for the hole that was closed, and then run that exploit against websites still running the old version — all without a person directing it.

Two months ago, in the previous edition of the same quarterly tracker, GTIG said it had not seen fully autonomous attack pipelines deployed in the wild. That has now changed, the new report says. Titled "Agentic Threat Activity Crosses a Significant Threshold", the report describes the adversarial AI landscape as having reached a critical inflection point.

What Google found on the exposed server

The staging server held more than the framework itself. GTIG found data taken from victims sitting on the same machine, alongside scripts for brute-forcing logins, installing webshells, and even removing webshells left behind by rival attackers. A configuration file carried Russian-language comments, including one that told whoever set the system up where to enter its language-model API keys. Google has not named the operator or attributed the framework to a country.

Victim telemetry suggests the operator was most interested in French websites connected to media, entertainment and education. The report does not give a success rate for the framework or a count of the sites it broke into, and it does not establish whether the autonomous operations were responsible for the stolen data found on the server.

One screenshot in the report shows the framework inspecting the Spectra WordPress plugin, version 2.19.25 — a plugin with more than a million active installations — specifically a permission check on its popup route. Google presents this as an example of the framework at work. It does not say a flaw in Spectra was exploited.

Why the patch window is shrinking

The finding matters because it attacks the routine every organisation relies on: a vendor publishes a fix, and users install it before attackers can take advantage. GTIG says language models can now "dramatically reduce the window between patch disclosure and weaponized exploitation". Because the source material for a patched flaw is finite — the fix itself shows what was wrong — the report argues attackers may find it more efficient to point language models at known, patched flaws, called n-days, than to hunt for unknown zero-days.

A case from late August shows how fast that window can already close. Two flaws in print management servers from PaperCut were chained together and exploited as zero-days. Security researchers and the security company Huntress used large language models to build working exploits within a day of disclosure. A campaign tracked by GreyNoise, believed to have relied on language models to build its exploits, began four days after disclosure and hit hundreds of organisations. One foothold in that campaign reached domain administrator within minutes, GreyNoise found.

Other signs of agentic attacks in the same report

Exploit_agent is not the only autonomous system in GTIG's quarterly report. Open-source agentic penetration testing frameworks are being used to automate discrete steps of live intrusions. During one intrusion against the United States aerospace sector, a threat actor was observed using a multi-agent framework called CAIRN. Another actor ran a customised platform built on PentAGI between late August and early October to scan a wide range of organisations in Asia for vulnerabilities.

The report also looks beyond hacking. Influence operations are now using models for far more than writing posts: a turnkey platform one commercial contractor was building with help from the Gemini API is designed to run more than 100,000 social media accounts through cloud phone farms. And a suspected North Korean group tracked as UNC6836 used fake recruiter profiles carrying Google's SynthID watermarks to approach cryptocurrency organisations, then rigged Anthropic's Claude Code to download its SOMBERMEME backdoor by altering configuration files in internal code repositories. Google said it disabled the accounts and assets tied to the Gemini misuse it found.

The limits — and what defenders should do

Google is careful about the limits of what it observed. These frameworks currently "trade stealth for velocity", the report says: they generate noisy traffic and leave code-level traces, so good logging and monitoring can still catch them. But the direction of travel is clear. GTIG expects attacks needing little to no human interaction within about three months, and sophisticated, fully autonomous agentic attacks within about six, according to BleepingComputer's account of the report.

For anyone running a website or a web application, the practical message is simple. The gap between a fix being published and an attack using it is getting shorter, and for internet-facing software it is now a same-day job. Patch promptly, prioritise critical updates, and watch logs for the noisy, automated probing these frameworks leave behind. The patch itself is no longer just a fix — in the hands of an autonomous agent, it is also the instruction manual for the attack.

Related reading: Innopark IT previously covered another fallout from AI infrastructure misuse in Notion AI chat transcripts exposed in the Fireworks AI breach.

Sources

  • SiliconANGLE — "Google says an AI framework turned security patches into exploits with no human in the loop", 16 November 2026
  • CSO Online — "Google spots the first hacker AI system that finds and exploits flaws autonomously", 17 November 2026
  • Times of India — "Google threat team finds first AI hacking without human intervention", 16 November 2026
  • AI Weekly — "Google flags 'exploit_agent' as first autonomous AI attack in wild", 16 November 2026
  • Google Threat Intelligence Group — AI Threat Tracker, "Agentic Threat Activity Crosses a Significant Threshold", 16 November 2026 (via the above reporting)

What's Your Reaction?

Like Like 0
Dislike Dislike 0
Love Love 0
Funny Funny 0
Angry Angry 0
Sad Sad 0
Wow Wow 0
Nahid Joy I’m Nahid Hasan Joy, a technology writer, web developer, and digital enthusiast with a strong interest in the ever-changing world of technology.