Sign in

How AI turns a software update into a hacker's guidebook

Google found an AI turning security patches into attacks with no human involved. Experts say automatic updates help, but much of the work falls on companies.

Published on: Nov 18, 2026, 09:54:11 IST
Share
Share via
  • facebook
  • twitter
  • linkedin
Copy link
  • copy link

Google says it has found an AI system that reads the security fixes that software makers publish and then attacks websites that have not yet patched the vulnerabilities, all without a human telling it what to do, leaving people who put off their updates with less time than they think.

For a cyber-attackers, a patch made public can double as a guidebook to machine that has not updated the software yet. (Reuters Photo)
For a cyber-attackers, a patch made public can double as a guidebook to machine that has not updated the software yet. (Reuters Photo)

The finding leads the latest AI threat tracker from Google Threat Intelligence Group, the company's team that follows malicious attacks on the internet, and Mandiant, its incident-response unit, released on Tuesday. Google's researchers say it is the first time they have seen attackers use AI to find software flaws and break in through them "without a human in the loop".

Only eight months ago, Mandiant was sounding a good deal calmer. In its annual M-Trends report, published on March 23, it wrote that it did not "consider 2025 to be the year where breaches were the direct result of AI".

Four days before Google's report, on November 13, the US Federal Bureau of Investigation (FBI) issued an alert about ARTEX, a free AI hacking kit that attackers have turned on organisations around the world, from banks to government departments.

Also read: Insider threats, network air gaps, secret rooms: The spycraft behind IAF chief's warning

A hacker's job, done by software

Google's threat hunters found a server someone had left open to the internet, and on it, software that was doing a hacker's work from start to finish on its own. Google calls it exploit_agent. It plugs into several large language models, the technology behind chatbots such as ChatGPT, including some running on the operator's own computers.

The system ran round the clock, comparing popular website software before and after security fixes were announced to work out what each fix had repaired. Then it wrote an attack and fired it at sites still running the old version, without waiting for anyone to sign off on it.

The server also held data stolen from victims, and scripts for what comes after a break-in. They could scout a victim's network and try password after password to get into accounts. One could plant a webshell, a hidden back door on a web server that lets an intruder come back whenever they like, and another could clear out webshells that rival hackers had left behind.

Records on the server suggest the operator was particularly interested in French websites in media and entertainment, and in education. Google has not said who was behind it.

Also read: AI is learning to protect itself and could resist being switched off: Stuart Russell

A game of ‘spot the difference’

Software makers fix security flaws by sending out updates, known as patches. Every patch is public, though, and anyone can put the new version of a program next to the old one and play spot the difference. The lines of code that change tend to point straight at the flaw that was just fixed. For an attacker, the update doubles as a guidebook to every machine that has not installed it yet.

Security researchers call these n-days. They are flaws that the maker knows about and has fixed, but that stay open on any machine where the update hasn't been installed. A zero-day, by contrast, is a flaw the maker hasn't discovered yet, which means there is no fix to install at all.

Google's report explains why attackers find AI more useful for the n-day hunt. Hunting for a zero-day means searching a program far and wide, whereas an AI going after an n-day has a much smaller pile of material to read, mainly the old and new versions of the code and the maker's notes on what changed.

Comparing patches is slow, labour-intensive work, and Google says attackers now hand it to AI, along with the job of turning what it finds into a working attack, and that is shrinking the "patch window", the time organisations have to install a fix before attackers strike.

A shrinking headstart

Disclosures of new flaws and the attacks built on them nearly doubled between January and August 2026, Google found. Zero-days rose far more slowly, from an average of 8 a month in 2025 to 11 a month this year, though there were 22 in August alone.

In late August, attackers were already exploiting two flaws in a system that offices use to manage their printers. Used together, the two flaws allowed an outsider who had no password to run their own commands on the server. Once the flaws were made public, security researchers used AI to build working attacks within a day.

Four days after the disclosure, a security firm whose findings Google cites saw an attacker hitting hundreds of organisations, probably using AI to build the attack and pick its targets. Within minutes of getting a first foothold inside a company, the attacker had taken over the account with the widest powers on its network.

The FBI looked at a different stage of an attack, the break-in itself, and reported a similar pace. Some ARTEX intrusions went from first picking a target to privileged access, the kind of control normally reserved for a system's administrators, in under three hours.

The second case involved flaws in PHP, a programming language widely used to build websites, and AI helped find them in the first place. Before the official fix came out on October 21, a video showing an attack and an early copy of the fixed code had already been published, which let researchers and attackers alike play spot the difference before the patch was even available.

There was also a flood of junk to deal with. In the days after the fix, hundreds of supposed attacks were posted online that did not actually work, most likely written with AI's help. Security teams had to test them to rule them out while watching for the real thing.

The update button is the defence

Google says the break-ins it describes still rely on stolen passwords and unpatched systems, the same weak spots attackers used before AI.

AI lets attackers hit those weak spots far faster and on a far bigger scale.

India's cyber-security agency, the Indian Computer Emergency Response Team (CERT-In), made the same point in its October 15 advisory on AI agents. "In the incidents observed, AI found and exploited these conventional weaknesses faster, more exhaustively and at greater scale than human attackers," it said. Its May blueprint on defending against AI-assisted attacks asked organisations to fix known, actively exploited flaws on systems facing the internet "within 12 hours where feasible".

Some machines never get the fix at all. Danfoss, which makes refrigeration controllers, said most of the French supermarket systems that failed in October's cyberattacks were older models that no longer received patches, which means their owners have no fix to wait for.

Google's evidence is about website software, but the same logic applies to any device that gets security updates, phones and laptops included, since a published fix tells attackers where to look on every machine that hasn't installed it. Switching on automatic updates is an easy solution in such instances, because the device installs the fix without the user having to remember to do it.

Also read: Highly classified network used by US’s National Security Agency down: Report

AI in the attacker's kit

Attackers who don't want to build their own tools needn't bother, since ready-made AI hacking kits are free to download. They were built for penetration testers, the specialists companies hire to attack their own systems and find the gaps. Google says criminals and suspected China-linked spies are now using them, in one case against targets that included the US aerospace sector.

ARTEX went up on GitHub, a site where programmers share code, in July. Its developer took it down on October 8 after reports that attackers were abusing it, saying it had been meant for authorised security testing, but copies, modified versions and a paid edition continue to circulate. Researchers counted more than 1,600 ARTEX control panels exposed on the internet in early November, the FBI said.

According to FBI, the kit lets attackers with “minimal technical expertise” run complex campaigns against several targets at once. In one, against financial institutions in South Korea, attackers took the personal information of about 67,000 people. In another, they turned a cloud backup service into a route into more than 2,250 of its customers' accounts.

ARTEX has no AI of its own and plugs into models made by US and Chinese companies. In the setups attackers ran, low-cost Chinese models were the most common, the FBI said, partly because they showed no visible guardrails once a request was framed as security testing.

CERT-In found criminals using such kits against Indian organisations, where the kits cut the time from break-in to damage "from days to a few hours".

The catch for the attackers

Google says these AI tools give up stealth in exchange for speed. Left to run unchecked, they generate traffic investigators can spot and leave recognisable traces in their code, which raises the odds of getting caught. Language models also make things up, and Google says attackers using them have had to contend with invented details and even targets that do not exist.

The FBI says ARTEX is noisy too, and generates plenty of traffic that may set off alarms. But its attacks move so fast that defenders have little time to organise a response before the attackers get what they came for.

Google expects the tools to grow quieter and need even less human help as the AI behind them improves.

  • Binayak Dasgupta
    ABOUT THE AUTHOR
    Binayak Dasgupta

    Binayak Dasgupta is a senior editor at Hindustan Times, New Delhi, where he has worked since 2013. He heads the paper's Page 1 team and is part of the editorial leadership overseeing daily news operations. He has previously led the national news desk. His work sits at the intersection of specialist depth and emerging public consequence — stories where complexity and nuance are in themselves the journalistic challenge. He shapes Hindustan Times's coverage and editorial positions on technology and artificial intelligence, information security, environment, health and science — steering these subjects from the commissioning of reporting to editing and the framing of the paper’s institutional view. His work spans analysis, commentary and reportage and in the past has led the paper's coverage of Covid-19 pandemic science. His interviews and hosted sessions include Nobel laureates, technology leaders, economists and climate justice scholars. In recent months, Binayak has overseen the coverage of the Air India Flight 171 crash and its investigation, India's energy exposure during the West Asia conflict, the trade dispute with the US, Delhi's air quality emergency and the evolving conversations around AI, including the IndiaAI Summit.Read More