Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarThe Decoder→ original

Google Catches an AI Agent That Hacks Websites on Its Own, With No Human in the Loop

Google Threat Intelligence Group has observed, for the first time, an attack in which AI finds a vulnerability and exploits it on its own, with no human in the chain. On an open server, researchers found an exploit_agent framework. It compares code before and after a patch, identifies the flaw, writes an exploit, and runs it. Google places this at Level 5 and expects more powerful attacks within months. That makes it urgent to shorten the gap between a patch's release and its installation.

AI-processed from The Decoder; edited by Hamidun News
Google Catches an AI Agent That Hacks Websites on Its Own, With No Human in the Loop
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Key metric #1Targets were French media, entertainment, and education websites.
Key metric #2Russian-language comments in the configuration files hint at a Russian-speaking operator.
Speed matters more than stealththese attacks leave traces, which makes them easier to spot.

Until now, AI in attackers' hands has been an assistant: drafting a phishing email, suggesting a piece of code, or analyzing someone else's vulnerability. Google says it has now seen the next step: a chain where no human is needed at all. Google's Threat Intelligence Group (GTIG), in its AI Threat Tracker report, described the first recorded case of fully automated vulnerability discovery and exploitation.

What They Found on the Server

GTIG discovered an open server running an AI framework called exploit_agent. It connects to several language models at once and carries out the entire cycle on its own.

The mechanics work like this. When a developer releases a patch, the code leaves a trail: you can compare the version before and after the fix and see exactly what was repaired. This is an old technique, known as patch diffing, and specialists used to do it by hand. The agent does the same thing automatically: it takes the code difference, determines which vulnerability was closed, develops an exploit, and runs it against sites that have not yet updated.

A simple analogy: a lock maker releases an update and says it fixed a weak spot. Normally, a burglar would need a long time to figure out which one. Here, the program compares the old lock with the new one, works out where the flaw was, and tries it on everyone who has not replaced their lock yet.

According to data from the server, the operator's targets were in France: media, entertainment, and educational sites. Comments in Russian turned up in the configuration files. Google cautiously says this points to a Russian-speaking operator, but it is an indirect indicator, not proof.

Why Google Calls This a Turning Point

Google rates the current activity at Level 5. This designation describes situations where AI tools carry out several stages of an attack with minimal human involvement. GTIG calls the development a critical inflection point and warns that more powerful attacks will appear within a matter of months.

The main problem is speed. For many companies, the window between a patch's release and its installation is measured in days or weeks. While analyzing a patch required an experienced researcher, that window was relatively safe for mass-scale attacks. If analysis and exploit writing take an agent minutes and require no human, every public patch becomes a signal to attack everyone who is running late.

A Reassuring Detail and Takeaways for Practitioners

There is also some good news. According to GTIG, such automated attacks currently sacrifice stealth for speed. The agent operates noisily and leaves conspicuous forensic traces, which makes it easier to detect. But that reflects the tool's current state, not a guarantee: operators could teach future versions to be quieter.

What this means for teams responsible for products and infrastructure. These are my conclusions from the mechanics described, not recommendations from the report:

  • Shorten the time it takes to install critical patches. If a week of testing used to be the norm, that timeline should be reconsidered for public services.
  • Do not leave old or forgotten services unattended. Automation does not pick targets by importance; it simply checks everything in sequence.
  • Strengthen logging and monitoring. Because these agents are noisy, anomalous spikes in requests and repeated exploitation attempts should be noticed and investigated quickly.
  • If you build AI agents with access to code and networks, think through the constraints: what the agent may run and where it may connect.

What the message leaves out matters too. Google does not disclose how many sites were actually breached or how successfully the agent found working exploits. The alarming signal is the fact itself: the full cycle, from patch analysis to exploit execution, has already been assembled and is operating on a real server. For defenders, this means preparing for an adversary that works at machine speed.

Story

Key Questions Answered

What They Found on the Server?

GTIG discovered an open server running an AI framework called exploit_agent. It connects to several language models at once and carries out the entire cycle on its own. The mechanics work like this. When a developer releases a patch, the code leaves a trail: you can compare the version before and after the fix and see exactly what was repaired. This is an old technique, known as patch diffing, and specialists used to do it by hand.

Why Google Calls This a Turning Point?

Google rates the current activity at Level 5. This designation describes situations where AI tools carry out several stages of an attack with minimal human involvement. GTIG calls the development a critical inflection point and warns that more powerful attacks will appear within a matter of months. The main problem is speed. For many companies, the window between a patch's release and its installation is measured in days or weeks.

What are the key details on A Reassuring Detail and Takeaways for Practitioners?

There is also some good news. According to GTIG, such automated attacks currently sacrifice stealth for speed. The agent operates noisily and leaves conspicuous forensic traces, which makes it easier to detect. But that reflects the tool's current state, not a guarantee: operators could teach future versions to be quieter. What this means for teams responsible for products and infrastructure.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…