Quick Summary
- Google says it found an AI tool that can spot a software flaw, write an exploit, and launch it with no human steering.
- The tool, called exploit_agent, studies what changed in a software patch to work out what was fixed.
- It goes after old bugs that already have a fix but were never installed on some systems.
- Its server also held stolen data and break-in scripts, mostly tied to French media, entertainment, and education sites.
- Other AI-driven attacks this year moved just as fast, from hours down to seconds.
- These tools are quick but noisy, so security teams can still spot and disrupt them.
- Fast updates and a few basic habits still cut most of the risk.
Autonomous AI attacks used to sound like a movie plot. Google’s Threat Intelligence Group now says they are real. Its researchers found a tool that can read a software update, figure out the bug it fixed, write code to abuse that bug, and run the code with no person involved. This article covers what was found, which attacks have already been reported, and what everyday teams can do about it.
An AI Hacking Kit Found on an Unlocked Server
Google’s latest AI Threat Tracker report describes a framework called exploit_agent. Researchers found it on an exposed staging server, which is a machine attackers use to test their tools. Google says this is the first quarter in which it saw attackers automate both the finding and the exploiting of flaws with no human in the loop.
The framework connects to several large language models. Some are leading commercial models. Others run on hardware the attacker controls. It uses them to keep checking popular web applications for weak spots.
Reading a Patch Like a Treasure Map
Whenever a company ships a security update, exploit_agent compares the code from before and after. The differences show where the bug was. It then drafts an exploit for that bug and runs it on its own. In plain terms, the tool reverse-engineers a fix and turns it into an attack.
Patches exist to protect people. They also hint at what was wrong in the first place. That hint is valuable to anyone targeting systems that have not installed the update yet.
Old Bugs Are Still Fair Game
Bugs that are public and already fixed, but still open on some systems, are called n-day vulnerabilities. A zero-day is different because it gets abused before any fix exists. This matters because exploit_agent is not necessarily finding brand new flaws. It is automating attacks on flaws that already have a fix.
That puts the spotlight on the gap between a patch coming out and a patch being installed. Every day of delay gives automated tools more time to find an open door.
The Stolen Data Sitting Next to the Tool
The server held more than the framework. Google also found data taken from victims, along with scripts for what happens after a break-in.
Those scripts handled internal reconnaissance and brute-forcing. Some installed webshells, which are small malicious scripts that give an attacker remote control of a web server. Others removed webshells left behind by rival attackers.
The victim data suggested the operator cared most about French media and entertainment sites, plus the education sector. Google has not said how many organizations were hit or who they were. It also has not named the operator or said that one specific exploit from this tool caused the stolen data.
Real Attacks That Already Happened
Exploit_agent did not appear out of nowhere. Here are recent incidents that researchers and officials have reported.
- exploit_agent (Google, November 2026): An exposed server held a fully automated exploit pipeline, stolen victim data, and post-break-in scripts. The activity leaned toward French media, entertainment, and education sites.
- PaperCut servers (late August to early September 2026): A security firm reported that attackers used hundreds of AI agents to compromise at least 440 PaperCut instances across 395 organizations in 48 countries. Credentials were taken from 280 victims. The attacker went from an empty workspace to a first real break-in in just under four hours. Once the wider campaign began, at least 11 organizations fell in 26 seconds.
- Cloud credential theft (September 2026): Google documented an attacker who, after breaking into a cloud resource, planned, built, and ran a credential-harvesting campaign in under six hours. The attacker used an AI coding chatbot and preconfigured agent instructions.
- About 100 company breaches (October 2026): A hacker used AI agents to break into about 100 companies. Researchers estimated the cost at roughly $5,000 in AI usage, which works out to about $50 per successful breach.
- ARTEX (FBI warning, November 13, 2026): The FBI says multiple threat actors have used an open-source AI penetration testing framework since at least August 2026. Some intrusions reached privileged access in under three hours. The targets included financial, government, technology, defense, and healthcare organizations. The advisory also points to an October supply-chain attack on a cloud backup product that exposed Microsoft 365 customer environments.
The trend is easy to see. AI first helped attackers with small tasks. Then AI agents took over larger chunks of the work. Exploit_agent is the first fully automated pipeline Google has documented from start to finish.
Can Anyone Slow Autonomous AI Attacks Down?
Yes, and Google says so itself. These tools currently trade stealth for speed. Running without limits leaves behind forensic traffic, recognizable identifiers, and code-level traces. That makes the tools easier for security teams to spot and disrupt.
There is another small relief. Growing AI use has created a flood of exploit claims that do not actually work. Security teams now spend extra time checking which claims are real.
The wider numbers still show pressure. From January to August 2026, vulnerability disclosures and exploits nearly doubled. Zero-day exploits averaged eight a month in 2025, rose to 11 a month in 2026, and hit 22 in August.
How to Keep Your Own Systems Safe
Google’s chief threat analyst points out that attackers are mostly using known weaknesses. What changed is the speed and scale. He argues that defenders will need automated, agent-based defenses of their own because manual effort cannot keep up.
You do not need a big budget to improve your odds. These habits line up with how the attacks above worked:
- Install updates quickly. Patches are the exact clue these tools read. The faster a fix goes in, the less time attackers have to use it.
- Fix the known-exploited bugs first. CISA recommends that all organizations review its catalog of known exploited vulnerabilities and prioritize those fixes.
- Know what faces the internet. The PaperCut attackers built their target lists with an internet scanning service. Fewer exposed systems means fewer targets.
- Watch for noisy traffic. Automated tools leave traces. Logging and alerts give you a chance to catch them early.
- Protect your logins. Credential theft was a goal in several of these attacks. Strong, unique passwords and multi-factor sign-in limit the damage.
- Check before you panic. Many AI-generated exploit claims do not work. Confirm a threat is real before reacting.
Final Thoughts
Autonomous AI attacks are no longer a distant risk. Google has documented a tool that reads patches, writes exploits, and fires them off with no human involved. Other incidents this year show the same pattern of speed.
The good news is that most of this still depends on unpatched and exposed systems. These tools are loud, and many AI exploit claims fall flat. Staying current on updates and watching for odd traffic remains the most practical defense.
Discover how AI is reshaping technology, business, and healthcare—without the hype.
Visit InfluenceOfAI.com for easy-to-understand insights, expert analysis, and real-world applications of artificial intelligence. From the latest tools to emerging trends, we help you navigate the AI landscape with clarity and confidence