1. A Google AI hacking tool discovery with no human at the controls
Google said on Monday, November 16, that its researchers had discovered a piece of attack software able to locate weaknesses in widely used web software and exploit them entirely on its own. The Google AI hacking tool finding, published in the third-quarter report of the Google Threat Intelligence Group (GTIG), marks the first time the company has seen attackers remove people from that process altogether.
The program, which Google has named exploit_agent, was not captured in the middle of an attack. Researchers came across it on a staging server, the kind of machine intruders use to store their tooling and get it ready before an operation. In this case the server had been left reachable from the open internet, which allowed GTIG to examine what was sitting on it.
What they found was a system wired into several large language models at once. Google said the tool connected to leading commercial models as well as to models the company describes as local, meaning the operator was not relying on a single provider to do the reasoning behind the attacks.
The discovery matters beyond the single server involved. Security teams have long assumed that turning a newly disclosed weakness into a usable attack takes skilled people and some amount of time. A tool that performs the whole sequence automatically challenges that assumption, and it does so at a moment when officials and economists are warning about what autonomous software agents could do to older systems.
2. How does exploit_agent work?
The tool waits for software makers to fix a flaw, works out what the fix reveals, and then attacks installations that have not yet applied it. According to Google, exploit_agent kept watch over popular web software for newly released patches. Each time one appeared, it compared the code as it stood before the fix with the code afterward in order to pinpoint the weakness the developers had just closed.
The method resembles laying two drafts of a document side by side to see exactly what an editor changed. A patch is, in effect, a public record of where a problem used to be. Anyone who can read the difference between the old and new versions learns where unpatched systems remain exposed.
Once the weakness had been identified, the tool moved to the next stage without waiting for instructions. It drafted an exploit, a piece of code built to take advantage of the flaw, and then ran that code automatically.
Comparing patched and unpatched code is an established technique among both defenders and attackers. What is different here is that the comparison, the writing of the attack code and its execution were all chained together and handed to language models, so the interval between a patch being published and an attack being attempted no longer depends on a person doing the work.
3. Why Google calls it a first
GTIG's report draws a clear line between this case and earlier uses of artificial intelligence by attackers. The group wrote that it had, for the first time, observed threat actors using large language model frameworks to fully automate the discovery and exploitation of vulnerabilities, with no human in the loop.
That phrase carries weight. Previous accounts of AI in intrusions have generally described models assisting people: helping to write a phishing message, speeding up research on a target, or suggesting code that an operator then reviews. In those cases a human still decides what to attack and when. With exploit_agent, Google says, those decisions were delegated to the software.
The report offers one concrete glimpse of the tool at work. A figure in the document shows it examining access controls in Spectra, a WordPress plugin also known as Ultimate Addons for Gutenberg, at version 2.19.25. Access controls are the rules that determine which users are allowed to perform which actions on a site, and errors in them can let an outsider do things reserved for administrators.
Google's report does not say whether that particular analysis ended in a successful attack. The example establishes what the tool was looking at, not what damage, if any, followed from it.
4. What else was sitting on the server
Alongside exploit_agent, the exposed machine held a broader collection of intrusion tooling. Google found scripts for internal reconnaissance, the work of mapping out a network once an attacker has made it inside. There were also scripts for brute-forcing passwords, which means trying large numbers of combinations until one works.
A third set of scripts was designed to plant webshells, hidden control panels left on a compromised site so that the intruder can return later without having to break in again. Together, the files describe a full sequence of activity: get in, look around, widen access and keep a way back.
The server also contained data taken from victims, which indicates the operator had gone beyond experimentation. Records on the machine suggest that the targets included French websites in the media, entertainment and education sectors.
Google's report does not put a number on the victims, and it does not attribute the operation to a named group or government.
5. Russian-language comments in the files
Notes written in Russian inside the tool's files provide the clearest hint about who built or ran it. Google published a screenshot of the comments and supplied translations of two of them.
The first reads: "LLM API keys for auto-exploitation (exploit_agent pilot)." The second is an instruction to whoever is setting the tool up: "Enter the keys yourself — right here, not in the chat. Both lines are required."
An API key works like a password that allows a piece of software to use an AI service. The comments therefore describe the place in the configuration where the operator supplies the credentials that let exploit_agent call on language models. The word "pilot" in the first comment suggests the authors regarded the tool as an early trial.
The language of code comments shows what its authors wrote in, and nothing more. Google's findings as described do not go further than noting the Russian-language text.
6. Zero-day attacks doubled as AI attack tools spread
Google places exploit_agent inside a wider acceleration in attacks on software flaws during 2026. The company describes disclosures and exploits as nearly doubling between January and August of this year.
The sharpest figures concern zero-day exploits, which are attacks on flaws the software maker does not yet know about and so has had no chance to fix. Google counted an average of 8 such exploits a month in 2025. That average rose to 11 a month in early 2026, and the count reached 22 in August.
The company also reported that openly available AI attack tools have moved from demonstration to practice. It named two open-source projects, CAIRN and PentAGI, and said it has seen both used in real intrusions.
For organizations that run common web software, the practical consequence is a shrinking margin for delay. A fix that has been published but not yet installed has always been a risk. A tool that reads the fix and attacks without human effort makes that waiting period more dangerous.
7. North Korean group used AI in crypto hacks
Digital asset companies feature directly in the report. Google said a North Korean group it tracks as MIDNIGHT NEPTUNE used AI in intrusions at crypto companies in North America and Europe.
The approach began with deception. Members of the group posed as recruiters on professional networking sites, a familiar route for reaching engineers who hold access to valuable systems. Google noted that the profile photos on those accounts carried SynthID watermarks, the markers Google embeds in AI-generated images, which showed the pictures had been produced by a model.
After gaining entry, the attackers altered a settings file in the companies' internal code repositories. The change caused a backdoor called SOMBERMEME to be downloaded whenever a developer opened Claude Code, an AI coding assistant. The routine act of starting a development tool thus became the trigger for installing the intruders' software.
The group also took steps to blend in. It created GitHub access tokens and named them after Microsoft Copilot, so that the activity tied to those tokens would look less suspicious to anyone reviewing it.
The disclosure comes as the industry absorbs a very large loss. Binance was recently hacked for roughly $6.3 billion, and blockchain forensics firms and the FBI attributed that exploit to hackers linked to North Korea. TRM Labs puts the total stolen in crypto by North Korea-linked entities at about $12.8 billion to date.
8. Warnings about autonomous agents are mounting
Concern about self-directed AI software is no longer limited to security researchers. Google's finding adds to a series of warnings that autonomous agents pose a threat to legacy systems built on the assumption that people, not programs, initiate activity.
Federal Reserve Governor Michael Barr said on Monday that such agents could trigger a bank run within minutes. In September, Torsten Slok, chief economist at Apollo, warned that AI assistants shifting household cash into higher-yielding accounts could drain banks of their cheap deposits.
Those warnings concern finance and Google's concerns software security, but they share a premise: processes that once moved at the pace of human decisions can now be carried out by software acting continuously and without supervision. In banking, that could mean deposits leaving faster than an institution can respond. In security, as exploit_agent shows, it means the step from a published fix to an attempted attack can be taken without anyone at a keyboard.
GTIG's report leaves several questions unanswered, including who operated the server and whether the WordPress plugin analysis led to a compromise. What it does establish, in Google's account, is that fully automated discovery and exploitation of software flaws has been observed in an attacker's hands for the first time.
