top of page

Google Finds Hackers Using AI Agents to Turn Patches Into Exploits

3 minutes ago
6 min read

Google's Threat Intelligence Group said on Monday, November 16, that attackers used an AI agent to turn software patches into working exploits with no human in the loop. Google called it a first. A published fix can now become an attack almost as fast as companies install it.


  • Google Threat Intelligence Group published its AI Threat Tracker report on November 16, 2026.

  • An exposed server hosted exploit_agent, a framework that turns patches into exploits automatically.

  • Zero-day exploits averaged 11 per month in 2026, then spiked to 22 in August.

  • Exploits for two PaperCut flaws appeared within a day of their disclosure in late August.

  • Stolen AI accounts ran up as much as $1 million in charges before cleanup.


Cover image with the headline Your Patch Is the Map over a dark server room with red warning light

What did Google find?


Google found a hacking framework that can read a security patch and write an attack against it. The Google Threat Intelligence Group, or GTIG, detailed its findings in a report released Monday by Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor. GTIG summed up the shift in one line:


“In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”

The tool, called exploit_agent, was found on an exposed staging server. It connects to a variety of frontier and local AI models. It compares popular web application code before and after a patch, identifies the flaw the patch fixes, and then drafts and runs an exploit.


The same server also hosted scripts for the steps after a break-in, such as internal reconnaissance, password cracking and installing webshells. Webshells are backdoors hidden on a web server that let an attacker return whenever they want. One script even removed webshells that rival attackers had left on the same machines.


The report draws on Mandiant incident response work, GTIG's tracking of threat groups and data from Google's own platforms. Beyond hacking tools, it covers propaganda campaigns that have fully built AI into their operations, and criminals who sell stolen access to AI models.


How does an AI agent turn a patch into an attack?


A patch points directly to the vulnerable code, which makes it a roadmap for attackers. The technique, known as differential analysis, compares the old code with the fixed code line by line. Human experts have used it for years, but the work is slow and can take days.


The agent does the same thing at machine speed, continuously and across many applications at once. GTIG published an example of it inspecting access controls in version 2.19.25 of Spectra, a popular WordPress plugin. Comments in the framework's code were written in Russian, with English translations in the report. Data on the server suggested the operator was targeting French media, entertainment and education sites.


The numbers already show the window shrinking. Zero-day exploits averaged 8 per month in 2025 and 11 per month in 2026, then jumped to 22 in August. Overall disclosures and exploits nearly doubled from January to August 2026.


The PaperCut case shows how fast it moves


Two flaws in PaperCut print management software show how quickly attacks now follow disclosure. In late August, attackers chained an authentication bypass, CVE-2026-81578, with an unsafe class loading flaw, CVE-2026-82078. Together they allowed remote code execution without a login.


Using AI models, researchers and the security firm Huntress built working exploits within a day of disclosure. Four days after disclosure, GreyNoise reported a campaign against hundreds of organizations. The attacker moved from initial access to domain administrator within minutes.


AI also creates extra work for defenders. Two PHP flaws, CVE-2026-109618 and CVE-2026-109619, were found with AI-assisted analysis. Hundreds of AI-generated proof-of-concept exploits followed, many of which did not work, and each one had to be checked.


Attack frameworks are going multi-agent


Exploit_agent is just one of several AI attack tools now in use. GTIG named ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. Most started out as open source penetration testing tools for defenders.


CAIRN is a set of specialized agents that share a common "blackboard" and work toward a goal, starting from as little as a target IP address. GTIG saw it used against targets including the U.S. aerospace sector. CyberStrikeAI connects AI models to more than 100 security tools, and HexStrike AI wraps more than 150.


Google's analysts added a warning about where this is heading:


“Threat actor adoption of these frameworks represents a paradigm shift towards intrusions occurring at machine speed, with the automated chaining of complex attack paths scaling operations significantly.”

For now, these tools have one weakness that defenders can use. According to GTIG, they trade stealth for speed, leaving noisy network traffic, recognizable markers and code traces that make them easier to spot. Google Security Operations has already released detection rules for ARTEX.


Stolen AI accounts are now a business


Attackers are also stealing companies' access to AI and reselling it. GTIG calls the practice LLMjacking. Attackers often start with valid credentials taken from public GitHub repositories, or with session tokens stolen through fake websites. Some fraudulent charges topped $100,000 in a single day, and the worst cases reached $1 million before cleanup.


In one case, a botnet using the same stolen credentials made millions of AI model calls in about 24 hours before the credentials were revoked. Stolen access is sold through underground storefronts that GTIG calls LLM mixers, at prices from 10 cents to $25 per million tokens. One attacker used a sandbox escape in LiteLLM, CVE-2026-40217, to turn limited AI keys into full admin credentials.


Attackers are also targeting developer tools. A North Korea-linked group tracked as MIDNIGHT NEPTUNE modified Claude Code settings files in a victim's code repositories. When developers launched the tool, it downloaded a backdoor called SOMBERMEME.


A server pulled out of a data center rack with its fans, heat sinks and drive bays exposed and a red warning light on the front

What does machine-speed hacking mean for Silicon Valley?


Silicon Valley companies now have less time between a patch and an attack than ever before. If an agent can turn a patch into an exploit within hours, a monthly patching cycle hands attackers weeks to work. That is especially hard for startups, which tend to have small security teams and fast-changing code.


The report's findings point to a few practical steps for engineering teams. Shorten the time it takes to install patches, and verify public proof-of-concept exploits before acting on them. Treat AI coding assistant settings files as sensitive code, and watch AI usage and cloud bills for sudden spikes.


The same technology also helps defenders. Google points to Big Sleep, an AI agent from Google DeepMind and Project Zero that finds unknown vulnerabilities, and CodeMender, which helps patch them. The contest is now between AI agents on both sides.


Some security experts say attackers have the upper hand for now. Matt O'Brien, CEO of the AI safety company SnowCrash Labs, told Sparknify's Silicon Valley Unplugged that AI-enabled offense may currently have an advantage over defense. An attacker needs to find only one path in, he said, while defenders have to close all of them.



GTIG expects the trend to continue. It judged it highly likely that attackers will keep using these frameworks against organizations around the world, given their low barrier to entry and ability to scale.


The harder question is how much to share. Open source penetration testing tools let security teams test their own systems, and a blanket ban would hurt the people who use them legitimately. Critics argue that releasing powerful attack agents publicly gives criminals a ready-made toolkit. Patch publishers face a similar tension, since detailed patches help users and attackers alike.


Five things will show which way this goes.


  1. Whether GTIG's next quarterly tracker finds more exploit pipelines with no human involved.

  2. Monthly zero-day counts for the rest of 2026, after August's spike to 22.

  3. How quickly software vendors shrink the time between a patch and customer installs.

  4. New detections for agent tools like ARTEX from Google Security Operations and other vendors.

  5. Billing alerts and key controls from AI providers to curb LLMjacking.


The patch once marked the end of a security problem, and now it can mark the start of the next attack.


Subscribe to Sparknify for tech news and the people behind it.


Questions & Answers


Q1: I run a small SaaS startup in San Jose. If an AI agent can turn our security patches into exploits within a day, how do we keep our customers' trust?


Patch speed now matters as much as the patch itself, since exploits for the PaperCut flaws appeared within a day of disclosure. Being open about fixes also builds confidence. A startup can tell that story to investors and decision makers through a package that includes a featured news article and a curated introduction, at sparknify.com/startup-pr-service.


Q2: What are banks like JPMorgan Chase and major websites doing to keep AI agents out of customer accounts?


Some are already blocking them. JPMorgan Chase and Charles Schwab have said they block AI agents from customer accounts on security grounds, part of a growing list of sites turning agents away. Google's report shows why that caution exists.


Q3: Hackers are stealing AI accounts now. Should I be concerned about connecting Google Gemini to my bank and card data?


Google says Gemini can read data in linked accounts but cannot move money. Still, attackers are stealing AI credentials and selling access, so it helps to know what Gemini can and cannot do with Google Wallet before linking any accounts.

Comments


Upcoming Events

No events at the moment

More Articles

Subscribe to Latest Sparknify News and Event Notifications

Current Role
bottom of page