Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →

Google finds AI agents that discover vulnerabilities and hack websites on their own

Google Threat Intelligence Group found an open server where AI agents ran an attack from start to finish without a human. The framework uses several language models: it compares code before and after a fix, finds the vulnerability, writes an exploit and launches it. The victims were French websites in media, entertainment and education. That is why patches should go out faster: the gap between a fix being released and an attack may now be very short.

AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Google finds AI agents that discover vulnerabilities and hack websites on their own
Source: TG: @ai_machinelearning_big_data (Machinelearning). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @ai_machinelearning_big_data (Machinelearning) ↗
Editor-in-Chief
Key metric #1Comments in the server's configuration files are written in Russian
Key metric #2The automation sacrifices stealth for speed and leaves traces that defenders can use
Key metric #3Google expects more powerful attacks of this type within a few months

Hacking a website used to require a human at every step: find a vulnerability, understand it, write the attack code, run it. Now Google says it has seen that entire chain carried out without a human. Specialists at Google Threat Intelligence Group discovered an open server running a framework of AI agents, and the company considers it a threshold moment in the history of cyber threats. For anyone who ships code to production, it is a reason to rethink their update habits.

How the automated attack works

The framework connects to several language models at once and hands out tasks to them. The first step is comparing versions of a program before and after a fix. When a developer releases a patch, the difference between the old and new code shows exactly where the hole was. A human researcher spends hours on that kind of analysis. An agent does it as routine.

Next, the system identifies the vulnerability itself, writes an exploit for it and launches it against targets. All of this happens without an operator. Put simply, it is a pipeline: a fresh patch goes in, and a working attack on anyone who hasn't updated yet comes out.

According to data from the server, the targets were French websites in media, entertainment and education. Comments in the configuration files are written in Russian. That is a detail from the find, but it does not by itself prove who is behind the infrastructure, and the original account does not specify attribution.

What is reassuring here

Google explicitly notes that for now the automation chooses speed over care. The agents act noisily and leave traces, which makes it easier for defenders to spot an attack. That is an important difference from targeted operations, where attackers stay undetected for months.

But Google isn't offering comfort: the company forecasts more powerful attacks within a few months. The logic is clear. Models are getting better at writing code, and frameworks are learning to hide. Today's noise is a feature of the first generation of such tools, not a permanent property.

What developers and product teams should do

First: shorten the time between a patch release and its installation. If an agent can automatically turn a fresh fix into a working exploit, the old cushion of "we'll update next week" is no longer safe. Critical dependencies and public-facing services should be updated first, and it is better to automate the update process.

Second: invest in detection. Since automated attacks are still noisy, logs, alerts on anomalous requests and rate limiting give you a real chance to stop them early. These are inexpensive measures, and they work best today.

Third: look at your services through the eyes of such an agent. Everything exposed to the outside, including forgotten test environments, open admin panels and old library versions, becomes a target faster than before. Taking inventory of your external attack surface is turning from a formality into practical hygiene.

Fourth, for founders and product managers: security is now tied to release speed. A team that ships its own fixes quickly and installs other people's fixes quickly is in a better position than a team with long approval cycles.

Limitations

Little is known. We know about one discovered server, the range of targets and the general principle of how the framework works. There is no data on how many attacks succeeded, which specific vulnerabilities were used or how well the agents perform without hints. Full technical details are published on the Google Threat Intelligence blog, and security teams should read those.

Still, the signal is clear. A fully automated chain of "patch, vulnerability, exploit, launch" has moved from theory to observed practice. Defenders will have to respond in kind: with automation in updates and monitoring.

Story

Key Questions Answered

How the automated attack works?

The framework connects to several language models at once and hands out tasks to them. The first step is comparing versions of a program before and after a fix. When a developer releases a patch, the difference between the old and new code shows exactly where the hole was. A human researcher spends hours on that kind of analysis. An agent does it as routine. Next, the system identifies the vulnerability itself, writes an exploit for it and launches it against targets. All of this happens without an operator.

What is reassuring here?

Google explicitly notes that for now the automation chooses speed over care. The agents act noisily and leave traces, which makes it easier for defenders to spot an attack. That is an important difference from targeted operations, where attackers stay undetected for months. But Google isn't offering comfort: the company forecasts more powerful attacks within a few months. The logic is clear. Models are getting better at writing code, and frameworks are learning to hide.

What developers and product teams should do?

First: shorten the time between a patch release and its installation. If an agent can automatically turn a fresh fix into a working exploit, the old cushion of "we'll update next week" is no longer safe. Critical dependencies and public-facing services should be updated first, and it is better to automate the update process. Second: invest in detection. Since automated attacks are still noisy, logs, alerts on anomalous requests and rate limiting give you a real chance to stop them early.

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…