Artificial intelligence wrote attack code based on a security patch, and even tried it out
While a company's IT staff are still arranging to install the latest security update, the attackers' artificial intelligence may already be examining how to exploit the very flaw that has just been fixed on machines that have not yet received the patch. Google's cybersecurity researchers have now documented for the first time an attack system that carried out the entire process — from comparing the patched code, through writing the attack code, to testing it — without human intervention, significantly shortening the time left for defenders.
According to a report published on Monday, November 16, by the Google Threat Intelligence Group (GTIG) and the company's Mandiant unit, attackers are increasingly entrusting tasks to artificial intelligence systems capable of carrying out instructions on their own. While in the past they mainly asked the models for coding help or used them to speed up individual steps of an attack, analysts have now found a system that links together and independently completes all the steps needed to find and exploit a vulnerability. Google considers this a turning point, since its previous report described only partially automated attack processes.
What happens when a program receives an update?
When a security flaw is found in a program — for example, in software that runs a website — developers release a fix for it, known in IT jargon as a patch. Installing this update can eliminate the vulnerability in question, but the release of the patch also gives attackers valuable information, because comparing the old and new versions of the program can reveal where the flaw was and how it could be abused. From this knowledge, an exploit can be created — the attack code used to break into systems that have not yet been updated — which means the release of a patch usually triggers a race between attackers and operators. Understanding the flaw and creating a working attack program could happen quickly even before, and automated tools already helped with this work, but more complex cases could require significant expertise and analysis lasting days or even weeks. During that time, companies and institutions had some time to install the update before anyone could exploit the security gap that remained open.
Google's analysts discovered a software package called exploit_agent on an inadequately protected server used to prepare attacks, consisting of several artificial intelligence agents working together. In this context, an agent means a program that uses tools to solve a given task, carries out actions, and then continues its work based on the results, making it capable of far more than simply answering questions in a chat window. The system that was found was connected to several large language models, including advanced commercial services and locally run models, and its job was to continuously compare the code of popular web applications before and after patching. It tried to identify possible vulnerabilities from the changes, created code designed to exploit them, and then automatically executed it as well, meaning its operator did not have to personally oversee every single attempt.
Besides data stolen from victims, programs were also found on the server that were used for further reconnaissance of already compromised systems, mass password guessing, and the installation of so-called web shells. A web shell is a program placed on a hacked website through which an attacker can remotely issue commands to the server and maintain access later on; among the seized programs was also one that removed similar tools belonging to other attackers, so that rivals could not use the same compromised system. Data found on the server relating to the victims suggested that the operator was particularly interested in French media, entertainment and education websites, and a configuration file contained comments in Russian explaining where to enter the access keys needed to use artificial intelligence services — one instruction specifically warned that the keys had to be inserted in the correct place in the file, not in the chat window.
Moreover, attackers do not have to write every program themselves, since they can also use freely available tools that were originally developed for security professionals to legally try to break into a company's systems and find weaknesses before criminals do. Google has already observed the tools called CAIRN and PentAGI being used in actual attacks, while in the case of CyberStrikeAI it has so far only reported experimentation. CAIRN was used in operations against American aerospace targets, while another attacker used a modified version of PentAGI from late August to early October to carry out regular reconnaissance and vulnerability scanning against a wide range of Asian organizations. Analysts also found traces of a self-developed, automated attack system belonging to a China-linked group, which they assess was likely intended for espionage. The role of artificial intelligence, however, does not end with the intrusion, since China-linked actors also use it to process and interpret stolen data, while North Korean attackers use it to build fake professional identities, among other things to deceive cryptocurrency companies.
The report also separately addresses an abuse known as LLMjacking, in which criminals obtain a company's access credentials for cloud services or artificial intelligence models and then use these services at the victim's expense. Among the cases examined by Mandiant was one in which more than one hundred thousand dollars in unauthorized charges were incurred in a single day, and in the most serious cases the bill rose to as much as one million dollars before the abuse was stopped. According to data examined by Google, the number of publicly disclosed vulnerabilities and exploits nearly doubled between January and August 2026, and in August alone twenty-two exploited vulnerabilities were classified as zero-days. This means that at the time of the attack no patch was yet available for them, so operators could not have eliminated the flaw simply by updating, even if they otherwise maintained their systems carefully.
In the case of exploit_agent, however, it is primarily already-patched flaws that pose a danger, because machines remain vulnerable in the period between the release of a patch and its actual installation. If attackers can produce usable code faster than before, even within hours, then a delayed update can pose a far greater risk than operators had previously accounted for. Automation, of course, does not make every attempt successful, and Google also warns that many of the attack codes generated by artificial intelligence and advertised as working are in fact useless. Checking these, however, also takes time away from defense specialists, who must filter out false alarms along with recognizing real threats. At present, defenders have a slight advantage because these systems often sacrifice stealth for speed, leaving behind large volumes of conspicuous network traffic and easily recognizable traces, which means that on a properly monitored network it can be easier to detect their activity than that of a patient, cautious human attacker — although this alone does not prevent them from causing damage.
Google expects that as the models and the programs built on them develop, less and less human involvement will be needed, while attacks may also become harder to detect. The company said it has taken action against accounts and projects abusing the Gemini service, and based on its findings has further strengthened its ability to detect abuse; it has also created behavior-based detection solutions for defenders, as well as a defense guide in cooperation with Mandiant. Delaying updates is thus becoming an increasingly risky habit, because the same patch that eliminates a vulnerability on one machine can also reveal to attackers where it is worth trying on others. While an operator waits for a more convenient time to install an update, the attackers' system may already be analyzing the changes, writing the code, and searching for servers where the security gap is still open.
Ugar
Did you like this article?
Support our work with a small donation
Secure payment via Stripe • Min. 2 EUR
Gesta recommendations:

Europe wants to replace Palantir, Musk showed why in Ukraine
European defense ministries want to reduce their dependence on key military tasks on the American company Palantir Technologies and other American technology firms, and an increasing number of European companies are competing for the market being vacated. Germany, France and the Netherlands are all building their own European systems for what Palantir's software has been used for or would have…

We receive the chemicals that disrupt our hormonal system even before we are born
Chemicals that disrupt the human hormonal system are now present almost everywhere in our environment: in plastics, cosmetics, toys and pesticides, and researchers have detected them in human blood, breast milk and umbilical cord blood. This means that exposure begins even before birth. This is revealed in a joint report by the United Nations Environment Programme (UNEP) and the World Health…

Oops: NSA's secret network goes down
A few days after Microsoft's worldwide outage and the flight ban that paralyzed American airports, service was also interrupted in one of the United States' most sensitive IT systems: the strictly classified internal network through which employees of the National Security Agency (NSA) share classified information went down. No connection has so far been established between the failures, and at…

Lasconi would support an alliance between the PNL and the USR, but fears that everyone will once again want too much
Elena Lasconi believes it would do Romania's right wing good if the PNL and the USR joined forces and tried to win back the trust of voters disillusioned with the parties under a new name, but she is less convinced that such an agreement could actually come about. The former USR leader explained her doubts by drawing on her own experience from her presidential campaign, recalling that in…

20 years in the EU: Negrescu calls for a new European plan for Romania
Twenty years after Romania and Bulgaria joined the European Union on January 1, 2007, Victor Negrescu, the PSD vice-president of the European Parliament, says Romania needs a new European programme that unites the political class and civil society. “We need a new European project for our country, a European plan for Romania,” he said at the European Parliament's building in Brussels, at the event…