Google: AI Hacking Tool Works With No Human

Neon cyber illustration of a robot hand typing code, with the words "NO HUMAN NEEDED."

For two years, security researchers have warned that artificial intelligence would eventually run a hacking campaign from end to end, with no human at the keyboard. On 16 November 2026, Google said it has now watched it happen. Its Google Threat Intelligence Group (GTIG) and Mandiant published a report describing an AI framework that reads security patches, works out what was fixed, writes an exploit and fires it — all on its own.

The framework is called exploit_agent, and Google calls its discovery a critical inflection point: the first time threat actors have used large language model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.

How exploit_agent works

The framework was found on a staging server the attackers had left exposed. It is a multi-agent system wired into several large language models, and it runs what Google calls continuous differential analysis on popular web application software.

The method is simple, and that is what makes it dangerous. When a vendor publishes a security patch, the framework compares the code before and after the fix. The difference between the two versions reveals what was wrong. The framework analyses that difference, drafts an attack for the hole the update closed, and then executes it automatically against websites still running the old version.

What Google actually found

The exposed server also held scripts for internal reconnaissance, brute-forcing passwords, installing webshells and removing other attackers' webshells, along with data taken from victims. Victim telemetry suggests the operator was most interested in French media, entertainment and education websites.

The configuration file carried Russian-language comments, including one Google translates as "LLM API keys for auto-exploitation (exploit_agent pilot)". The Register describes the operator as suspected Russian-speaking. Google itself names no group or country, and the report gives no success rate and no count of sites broken into.

Read one detail carefully: Google's screenshot shows the framework inspecting the WordPress plugin Spectra, version 2.19.25, which has over a million active installations. Google does not say the agent found a flaw in Spectra — it is an example of the popular software the framework watches, not a confirmed victim.

It is not just one tool

Exploit_agent is the headline, but the report describes a wider shift. Google says adversaries are now deploying open-source, multi-agent penetration testing frameworks in live intrusions, including CAIRN, PentAGI, CyberStrikeAI and ARTEX, with experimentation around HexStrike AI.

GTIG says it independently observed CAIRN used against targets including the US aerospace sector, and tracked a platform built on PentAGI conducting reconnaissance and vulnerability scanning against organisations across Asia between late August and early October 2026.

The numbers are stark: overall vulnerability disclosures and exploits nearly doubled between January and August 2026, and zero-day exploitation rose from 8 a month in 2025 to 11 a month in 2026, spiking to 22 in August. Two months ago, GTIG said it had not seen a fully autonomous attack pipeline in the wild. That has now changed.

Why the patch window is the real target

The attacks target n-days — flaws that already have a patch. That flips the usual security advice on its head. The dangerous moment is not the day a vulnerability is discovered; it is the days after the fix is published, when some systems are updated and others are not. Every unpatched server becomes a target the agent already knows how to hit.

What to do about it

For anyone running a website or a server, the practical lesson is blunt: patch promptly, and prioritise internet-facing web software first. The window between a patch announcement and an automated exploit is now measured in hours, not weeks.

That means keeping an inventory of what you run, turning on automatic updates where safe, and treating a patch announcement as the start of a countdown, not a routine chore.

Key takeaways

  • Google's Threat Intelligence Group and Mandiant reported on 16 November 2026 the first fully autonomous AI attack framework seen in the wild, called exploit_agent.
  • Exploit_agent compares web software code before and after a patch, identifies the flaw the patch closed, drafts an exploit and executes it with no human in the loop.
  • It was found on an exposed staging server alongside exfiltrated data; victim telemetry points to French media, entertainment and education websites.
  • Google does not name the operator; configuration comments were in Russian, and The Register describes the operator as suspected Russian-speaking.
  • Other frameworks — CAIRN, PentAGI, CyberStrikeAI, ARTEX — are also being used in live intrusions, and overall vulnerability disclosures nearly doubled between January and August 2026.
  • The attacks target n-days, flaws that already have a patch, so the time between a patch announcement and installation is now the main exposure.

Conclusion

Exploit_agent is not a superweapon — Google gives no success rate and no victim count. But it is a threshold: once an AI can turn a patch into an exploit on its own, every unpatched system is in a race against an automated reader that never sleeps. The defence has not changed, only the deadline: patch fast.

FAQ

What is exploit_agent?

Exploit_agent is a multi-agent AI framework found by Google's Threat Intelligence Group on an exposed staging server. It watches popular web software for patches, compares the code before and after each fix, drafts an exploit and runs it automatically, with no human in the loop.

Did exploit_agent hack a WordPress plugin?

Google's screenshot shows the framework inspecting the Spectra WordPress plugin, version 2.19.25, which has more than a million active installations. Google does not say the agent found a flaw in Spectra — the plugin is shown as an example of the popular software the framework watches.

What should website owners do?

Patch promptly, starting with internet-facing web software and the most widely used plugins. Keep an inventory of what you run, use automatic updates where safe, and treat a patch announcement as the start of a countdown — autonomous tools can now turn a published fix into a working exploit within hours.

Sources

Researched with AI assistance and reviewed for accuracy.

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for one tech and AI update every day: Daily Tech & AI WhatsApp channel.

 

Comments