Technology•

Google Finds an AI Framework That Turned Security Patches Into Attacks With No Human Involved

Google's threat researchers say they have seen attackers fully automate the discovery and exploitation of software flaws for the first time, using a multi-agent framework called exploit_agent.

Written By :MINRK
Google Finds an AI Framework That Turned Security Patches Into Attacks With No Human Involved

Photo by Shutter Speed on Unsplash

1. Google AI cyberattack report documents a first

Google's Threat Intelligence Group (GTIG) said on November 16, 2026, that it has observed attackers fully automating the discovery and exploitation of software vulnerabilities for the first time, with no person involved in the process. The finding, published in the group's latest AI Threat Tracker, centers on a multi-agent framework named exploit_agent that researchers recovered from an attacker's exposed staging server.

The report marks a change in how the company describes the role of artificial intelligence in hacking. Until now, its researchers had mostly recorded cases in which AI models helped human operators with individual tasks. In this instance, GTIG says the software itself carried the work from start to finish: it looked for weaknesses in widely used web applications, wrote code capable of abusing them and did so without anyone steering each step.

For security teams, the practical meaning is that a scenario long discussed as a future risk is now part of the record of real intrusions. The server that held the framework also contained data taken from victims, which investigators treat as a sign that the tooling had been used against live targets and was not simply an experiment sitting on a test machine.

2. How does exploit_agent work?

exploit_agent is a multi-agent framework that connects to several large language models and uses them to compare the code of popular web applications before and after a security patch, then builds a working exploit from what it learns. GTIG says the framework handles that entire sequence on its own and keeps repeating it, with no human supervising the individual stages.

The term multi-agent describes a design in which several AI components are each given a narrower job and pass their results to one another. One component might read source code, another might reason about what a change implies, and another might write and test attack code. Chaining the pieces together lets a system finish a long, multi-step task that a single prompt to a single model would not complete reliably.

The framework did not depend on one model. According to GTIG, it reached out to multiple large language models by itself, drawing on them as needed while it studied the software it was pointed at. That arrangement gives an operator flexibility, because the work is not tied to the availability or the restrictions of any single AI provider.

The targets of the analysis were web applications in common use. Software of that kind runs on a very large number of internet-facing servers, so a single working exploit can be reused against many organizations that share the same unpatched version.

3. Why comparing patched and unpatched code matters

Reading a fix to learn about the flaw it closes is one of the oldest techniques in security research. When a vendor releases a patch, the difference between the old and new versions of the code shows exactly which lines were changed. A skilled analyst can work backward from that change to understand the underlying weakness and, with enough effort, write code that triggers it on systems that have not yet installed the update.

GTIG acknowledges that researchers have relied on this approach for years, so the method itself is not what worries the group. What is new is that exploit_agent performs the whole cycle continuously and unattended. The steps that once required a specialist's time, from spotting the relevant change to producing a functioning exploit, are run by software that does not need rest and can move on to the next patch as soon as it finishes the last one.

That matters because of how patching works in practice. A security update protects only the organizations that apply it, and many take days or weeks to do so because updates must be tested and scheduled. The period between the release of a fix and its installation has always been a window of exposure. Automation on the attacker's side shortens the time needed to take advantage of that window, and GTIG warns that operations run this way can move quickly.

4. Who was targeted?

The activity tied to exploit_agent was aimed mainly at French websites in three sectors: media, entertainment and education. GTIG drew that conclusion from the material stored on the staging server where the framework was found.

That server offered an unusually direct view of the operation. Alongside the framework itself, investigators found data stolen from victims. The presence of that data is what separates the case from a proof of concept, since it indicates that systems had been compromised and information removed from them.

The server also held tooling meant for use after an initial break-in, including brute-forcing scripts, which try large numbers of passwords or keys against a login until one works. Tools of that type are typically used to widen access once an attacker has a first foothold. GTIG reads the overall setup as evidence of an organized operation at an advanced stage of attack automation, in which the AI framework was one part of a larger kit.

5. From assistant to operator

The significance of the finding lies in the contrast with what Google's researchers had documented before. Their earlier reporting during 2026 described attackers folding AI into their workflows as a helper. Models were used to speed up tasks such as credential harvesting, the collection of usernames and passwords, but people still directed the campaigns and made the decisions.

In that earlier picture, AI functioned as a productivity tool. It made existing operators faster or let less experienced ones attempt work that would otherwise have been beyond them, yet a human remained responsible for choosing targets, interpreting results and deciding what to do next. The technology supported the attack; it did not run it.

The exploit_agent case moves past that line. GTIG says that in the past quarter it identified threat actors using large language model frameworks to carry out vulnerability discovery and exploitation without a human in the loop, something it had not seen previously. The distinction is one of role: the software is no longer advising an operator but acting as one.

The claim is a statement about what Google's own investigations have turned up. It establishes that the capability exists outside laboratories and has been put to use, which is the threshold that defenders and policymakers have been watching for.

6. What it means for companies that run on web software

Any business whose services are delivered through web applications has a stake in the finding, and that includes much of the financial and digital asset industry. Exchanges, brokerages, payment firms and wallet providers all operate internet-facing software and rely on third-party components that receive regular security updates. The technique described in the report does not depend on the victim's industry; it depends on whether a known flaw has been left unpatched.

The traditional assumption behind many patching schedules is that turning a published fix into a usable attack takes meaningful skill and time, which gives organizations a grace period. A framework that performs that conversion automatically weakens the assumption. Security teams that treat internet-facing updates as routine maintenance may need to treat them as urgent, particularly for widely deployed web applications of the kind exploit_agent was built to study.

Scale is the other consideration. A human team can study only so many patches and pursue only so many targets at once, which has historically forced attackers to prioritize. Software that reads code changes and drafts exploits unattended is not bound by that limit in the same way. Flaws in less prominent applications, which might once have been ignored because they were not worth a specialist's time, become cheaper to pursue when the analysis costs little more than computing power and access to language models.

The staging server points to a further lesson about what follows a successful exploit. The framework sat beside brute-forcing scripts and stolen data, which suggests that the automated break-in was an opening move and that conventional tools were on hand to extend access afterward. Controls that limit what an intruder can do after getting in, such as strong and unique credentials and restricted administrative access, remain relevant even when the first step of an attack has been automated.

The report also has implications for how intrusions are investigated. When a person runs an attack, the pace and pattern of activity reflect human habits. An automated system working continuously behaves differently, and defenders who look for familiar patterns may need to adjust what they consider normal.

7. What comes next

GTIG does not present the case as a one-off. The group expects the trend to continue and has sketched how it thinks the threat will develop over the coming months. In its assessment, simple attacks that need little to no human interaction will become feasible first, and more sophisticated, fully autonomous operations are likely to follow.

That sequence reflects the usual way offensive techniques spread. A capability first appears in the hands of a well-organized group, as the tooling on the staging server suggests was the case here, and then becomes easier to obtain and operate as frameworks mature and are shared. The first wave of imitators tends to be unsophisticated, relying on the automation to make up for limited skill.

For now, the published findings give defenders a concrete example to plan around: a named framework, a description of how it turns patches into exploits, and a set of victims in French media, entertainment and education. The next editions of Google's AI Threat Tracker will show whether exploit_agent remains an isolated discovery or the first entry in a longer list.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY