Get a free audit
Apple credit Apple credits the researcher who runs our AI penetration testing.

Field Notes / AI Security

Google found an AI that reads security patches and fires the exploit itself. Nobody has to press go.

Google’s threat intelligence team has caught the first attacker framework that compares old and new code after a patch, writes the exploit and runs it with no human in the loop. What it means for how fast you patch, in plain numbers.

Author
Red Team Partners
Read
6 MIN READ
Filed
17 Nov 2026
Rows of server racks lit in a dark data centre.
22 zero-days in a month
Exploited zero-days in August 2026, against a 2025 average of 8 a month. Disclosures and exploits nearly doubled between January and August.

Google Threat Intelligence Group, 16 Nov 2026

Google’s threat intelligence team has found an attacker tool that does the work of a skilled exploit developer on its own. It reads a vendor’s security patch, compares the old code with the new, works out what was broken, writes the exploit and runs it, with no human in the loop GTIG . Google calls it the first time it has seen attackers fully automate both the research and the attack. For you, it means one thing: the day a fix ships for something on the internet is the day to install it.

What Google found

The framework is called exploit_agent, and it turned up on an exposed staging server wired into several commercial and locally hosted language models GTIG . Its job is to watch popular web application software, compare each release with the one before it whenever a patch is announced, and turn the difference into a working attack.

The same server held data already stolen from victims, plus scripts for brute-forcing logins, planting webshells and deleting webshells left by rival attackers. Victim data on the server points at French websites in media, entertainment and education. One comment in its configuration, written in Russian, tells the operator to enter the API keys “right here, not in the chat.”

Two months earlier, Google’s previous tracker said it had not seen fully autonomous attack pipelines used against real targets SiliconANGLE . That has now changed.

Read what the report leaves out, too. It names no product that exploit_agent attacked, gives no victim count and no success rate, and does not show that every exploit it wrote worked Hackread . A Russian comment does not tell you who runs it either.

exploit_agent is not alone. Google also saw open-source offensive agent frameworks used in live intrusions: CAIRN against targets including the US aerospace sector, and a customised PentAGI platform scanning organisations across Asia from late August to early October 2026 GTIG . We wrote about what CAIRN looks for on a website in our Cairn note.

How short the window has become

Google’s reasoning is simple. Hunting a zero-day means searching everything. A patch tells the attacker exactly where to look, so pointing a language model at fresh fixes is the cheaper bet GTIG . The fix itself is the map.

Case (2026)From public to exploited
PaperCut print serversResearchers and Huntress built working exploits with language models within a day of disclosure. GreyNoise saw a campaign four days after disclosure hit hundreds of organisations, with one foothold reaching domain admin within minutes GTIG .
PHP2Shell (PHP)Early patched code and a proof-of-concept video went public before the official patch on 21 October. Hundreds of fake, non-working exploits followed in the first days GTIG . Our PHP2Shell note has the detail.
N-able N-centralAbout five hours from exploit code in a public repository to the first server running a miner as root. Our N-central note has the timeline.

Those fake exploits are a cost of their own. Every claimed proof of concept has to be checked before anyone can say whether a system is exposed, and someone on your team spends the morning doing it.

How fast to patch, by system

Put the speed where the exposure is. You can’t patch everything in hours, and you don’t need to. This is the split we recommend.

SystemPatch withinIf you cannot
VPNs, firewalls, remote access, RMM consolesSame day the fix shipsTake it off the open internet or allow-list it until patched
Public web apps and their plug-insSame dayBlock the affected path at the web application firewall
Email and collaboration servers you host24 hoursRestrict external access and watch the logs
Internal servers and laptopsYour normal cycle, at most 14 daysNote the exception and who accepted it

Two rules sit under the table. A patch closes the door but does not remove anyone who came in before it, so check every internet-facing system for new accounts, webshells and odd scheduled tasks after you patch Hackread . And you can only patch what you know you expose, and the systems people forget are the ones attackers find first: a forgotten test server, an old admin panel, a supplier’s remote access box.

That inventory is the first thing a penetration test gives you, and if you run AI agents of your own, our AI red teaming shows what an attacker’s agent would do with them. Start with the list of what faces the internet. The patches are easier once you have it.

Questions readers ask

What is exploit_agent?

exploit_agent is an attacker framework that Google Threat Intelligence Group found on an exposed staging server and described on 16 November 2026. It connects several large language models, compares the code of popular web applications before and after each patch, works out the flaw the patch fixed, writes an exploit and runs it with no human directing each step. Google calls it the first time it has seen attackers fully automate vulnerability research and exploitation.

How fast do attackers exploit a vulnerability after a patch in 2026?

In the cases Google cites, days or hours. Researchers built working exploits for two PaperCut flaws within a day of disclosure, and GreyNoise saw a campaign against hundreds of organisations four days after disclosure. In November 2026 the N-central 4EverAble2Shell chain reached root on live servers about five hours after the exploit code went public.

Should we still patch monthly?

Not for anything reachable from the internet. Keep a monthly cycle for internal systems if it works for you, but patch internet-facing servers, VPNs, remote management tools and web applications the day a security fix ships. If you cannot patch that day, restrict who can reach the system until you can, and check it for signs of compromise once it is patched.

Does AI make attacks harder to detect?

For now it makes many of them easier to detect. Google says these frameworks trade stealth for speed: they generate noisy traffic, use recognisable identifiers and leave code-level traces. That favours teams who watch their logs, and Google expects the tools to get quieter as the models improve.

References

Sources

  1. Blum, A., Vanderlee, K., Xia, J. and Mainor, D. “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold.” Google Threat Intelligence Group, Google Cloud blog, 16 November 2026. cloud.google.com
  2. “Google says an AI framework turned security patches into exploits with no human in the loop.” SiliconANGLE, 16 November 2026. siliconangle.com
  3. “Google Spots AI Agent Turning Security Patches Into Exploits.” Hackread, 16 November 2026. hackread.com