Autonomous AI Cyberattacks: Google Finds exploit_agent Turning Software Patches Into Exploits With No Human in the Loop, and Why Patch Speed Now Matters Most
Google's threat intelligence team says attackers used AI frameworks to find flaws, write exploits and run them without a human in the loop for the first time. What is confirmed, what is not known, and what to do this week.
Where France, United States · Asia, Europe, Western Europe, North America, United States & Canada
On this page 10 sections
Google's threat researchers say attackers have started using AI systems that find software flaws, write the exploit and launch it with no person involved. If your business runs a website on popular software, the gap between a security fix going public and an automated attack trying it is getting shorter, so slow patching is now one of the biggest risks you carry.
What Google reported about autonomous AI attacks
Google Threat Intelligence Group (GTIG) and Mandiant published a report titled "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold". In Google's words, the headline finding is: "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."
LLM stands for large language model, the technology behind AI chat assistants. An agentic or multi-agent framework is software that has several of these models work through a task together, each handling a step, with little or no human direction.
SecurityWeek covered the report under the headline "AI Framework Autonomously Weaponized Flaws in Real Attacks, Google Says". This article works from Google's own report.
What is exploit_agent?
The main example in the report is a framework GTIG dubbed exploit_agent. Google describes it as "a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates with multiple frontier and local LLMs." GTIG says it found the framework on an exposed staging server.
How exploit_agent works
According to Google, the framework is set up to:
- Run continuous differential analysis of popular web application software.
- Compare the code before and after patch announcements.
- Analyze what changed to identify potential vulnerabilities.
- Draft exploits.
- Execute them automatically, without human intervention.
In plain terms: when a vendor ships a fix, the fix shows where the weakness was. This framework reads the before and after code, works out what was broken, builds an attack and runs it. Google's point is that the whole chain now runs with no human in the loop.
Who exploit_agent targeted
Google says the operator "was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector."
What is confirmed and what is not known
| Point | Status |
|---|---|
| Threat actors used LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop | Stated by GTIG in its report |
| exploit_agent is a multi-agent framework using multiple frontier and local LLMs, found on an exposed staging server | Stated by GTIG |
| The operator was interested in French media, entertainment and education websites | Stated by GTIG |
| Which web application software exploit_agent analyzed | Not named in the report material we reviewed |
| Which CVEs were exploited | Not named |
| How many websites were compromised | Not known |
| Who the operator is | Not named |
Because the software is not named, there is no single fixed version to install for this story. The defense is general: patch everything you run on the web, and do it quickly.
Other AI attack frameworks Google named
exploit_agent was not the only tool in the report. GTIG also named ARTEX, CAIRN, CyberStrikeAI, PentAGI, HexStrike AI, ZeroHollow and Agent Zero. Two come with extra detail:
- CAIRN was used against targets in the US aerospace sector.
- PentAGI: between late August and early October 2026, GTIG tracked a threat actor using a customized attack platform built off the open source version of PentAGI. Google says it was used for systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia.
The zero-day numbers in Google's report
Google says overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits, meaning attacks on flaws with no fix yet, rose only marginally: from a 2025 average of 8 per month to 11 per month in 2026, before spiking to 22 in August.
The report material we reviewed does not say the August spike was caused by AI frameworks, so we are not making that link.
Fast but noisy: what this means for detection
Google's framing is blunt: "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."
There is one piece of good news. Google says these frameworks "currently trade stealth for velocity." They move fast, but they are not quiet, which gives logging and monitoring a real chance of catching them.
That window may not last. Google says: "As the sophistication of these tools and underlying LLMs increases, we anticipate not only reduced human involvement in these operations but also increased stealth and sophistication of the activity."
Why this matters for a business website or company network
A tool that watches popular web software for new patches does not need to know who you are. If your site runs that software and the fix is not installed, your site matches what it is looking for.
Many businesses update their website once a month, or whenever the developer gets to it. Google describes attackers working at machine speed, so a monthly update cycle is a long time to stay exposed. We wrote about who is actually responsible for this in Who Is Patching the Software That Runs Your Business?, and about why an update notice is not proof of a fix in Updated Is Not the Same as Patched.
The exposed staging server is a lesson too. Test and staging copies of websites are often older, unpatched and forgotten. If you self-host, our self-hosting and infrastructure page covers how we keep those environments locked down.
What to do this week
| Action | How to check |
|---|---|
| List every piece of web software you run: content system, plugins, themes, store, booking or learning platforms | Ask your developer or host for a written list with current version numbers |
| Subscribe to each vendor's security advisories or release notes | Confirm someone on your team actually receives and reads them |
| Set a patch deadline measured in days, not weeks | Agree the target in writing with whoever maintains the site |
| Confirm fixes are live, not just downloaded | Check the version shown in the admin panel after each update |
| Remove plugins, themes and apps you no longer use | Anything not installed cannot be exploited |
| Take staging and test copies offline, or put them behind a login | Search for old subdomains such as test or staging that still load |
| Watch for bursts of unusual requests | Review web server or firewall logs; Google says these tools are currently noisy |
| Keep a recent backup stored away from the website | Restore one file from it to prove it works |
Autonomous AI attack questions people are asking
What is an autonomous AI cyberattack?
An attack where AI software handles the steps a human attacker used to do: finding the flaw, writing the exploit and running it. Google says it has now seen this done without a human in the loop.
What is exploit_agent?
A multi-agent AI framework, named by Google, that compares web software code before and after patches, finds the weakness, drafts an exploit and runs it automatically.
Was my website attacked by exploit_agent?
Not known. Google says the operator was interested in French media, entertainment and education websites. The report material we reviewed does not list victims or how many sites were hit.
Which software or CVEs did exploit_agent use?
Google's report, as we reviewed it, says popular web application software but does not name products or CVE numbers.
Does this affect businesses outside France?
The named targeting is French. The method, reading public patches for popular web software, does not depend on location, so fast patching matters for any business with a website.
Can AI-driven attacks be detected?
Yes, for now. Google says these frameworks currently trade stealth for velocity, so good logging and monitoring can spot them.
Will AI attacks get harder to spot?
Google expects so. It anticipates less human involvement and more stealth as the tools and models improve.
The bigger lesson
Every security patch is now read by machines as well as defenders. Once a fix is public, the clock starts, and the businesses that install it in days instead of weeks are the ones an automated tool passes by. Patch speed, a short list of what you run, and logs someone actually reads are the controls that matter most here.
Want help tightening this up? See security hardening or request a quote.
Get security alerts by email, plus our free checklist, at /security-alerts.
Sources: Google Threat Intelligence Group: GTIG AI Threat Tracker; SecurityWeek.
Have a question about this article?
Ask it here and we will answer you directly by email. Questions are private, never posted.
Security alerts by email
Hear about the next one first.
When attackers start using a flaw in software businesses run on, we email what it is, whether it touches you and what to do that week, with a round-up of our latest security writing. Free, one-click unsubscribe. You get the checklist straight away: 12 things to secure this week.
Want a result like this?
Security alerts, breaches and the technology a business runs on: what happened, who it affects and what to do.
Get a Tech QuoteKeep reading
Automated AI Cyberattack: Google Finds exploit_agent Turning Software Patches Into Live Attacks With No Human in the Loop, and What to Do This Week
Google says attackers used an AI framework called exploit_agent to find and exploit flaws in popular web softw...
Uber Data Breach: Hackers Stole Drivers' License Details From AI Vendor Fireworks AI, What Is Confirmed and What Businesses Should Check This Week
Uber says it notified 149,346 people after hackers stole drivers' names, addresses, dates of birth and license...
Cairn AI Framework Attacks: Unit 42 Says Attackers Paired DeepSeek and Zhipu GLM With an Open-Source Pentest Tool, Reached 10 of 19 Organizations, and What to Check This Week
Unit 42 says Chinese-speaking attackers paired the open-source Cairn framework with DeepSeek and Zhipu GLM, ta...