Tuesday, 17 November 2026

What is an AI cyberattack? Google found one running itself

Google caught an AI system that read security fixes, built attacks and launched them with no human help. Here's what that means, and why updating fast matters more now.

Dark server room with blinking racks and a laptop on a cart

The short version

  • Google Threat Intelligence Group says attackers used an AI system that found and exploited flaws with no human involved, a first.
  • The tool, called exploit_agent, compared old and new versions of web software to spot what each security fix repaired, then attacked unpatched sites.
  • Google expects simple attacks needing almost no human input within about three months, and fully autonomous advanced attacks within about six.
  • The best protection for ordinary users is still fast updates, two-factor sign-in and caution with unexpected job offers and files.
Quick read · 2 min

Google's threat research team says it has caught the first AI cyberattack system that worked with no human in the loop. Called exploit_agent, it compared old and new versions of popular web software after each security update, worked out which flaw the fix closed, then wrote and launched an attack on sites that hadn't updated.

Why it matters: the gap between a fix being released and criminals abusing the flaw is shrinking fast. In one recent case, researchers built working attacks within a day of disclosure, and a campaign hit hundreds of organizations days later.

  • Google expects simple, mostly automated attacks within about three months.
  • It expects advanced, fully self-running attacks within about six months.
  • Fake recruiters, AI phishing kits and stolen AI accounts are also on the rise.

What to do now: switch on automatic updates for every device, including your router, and use passkeys or two-factor sign-in on important accounts.

What is an AI cyberattack? It’s a break-in where artificial intelligence does the hacker’s work, and Google says it has now caught one where no human touched the controls. A criminal’s AI system read fresh security fixes for popular website software, figured out which hole each fix was plugging, wrote an attack for it and fired that attack off by itself.

That matters even if you never run a server. The time between a company shipping a fix and criminals abusing the flaw it closes has always been your window to update. Google Threat Intelligence Group, the company’s in-house threat research team, says AI tools are squeezing that window hard, and it expects simple attacks that need almost no human input within roughly three months.

01

What is an AI cyberattack, in plain terms?

An AI cyberattack is any hack where a large language model (the kind of AI behind chatbots) carries out part or all of the job. Until recently that mostly meant help around the edges: a more convincing phishing email, a translated lure, a bit of debugging on malware. A person still picked the target and pressed the buttons.

The new step is what security people call agentic attacks. Instead of one chatbot answering one question, several AI agents split the work. One scouts targets, another hunts for weak spots, a third writes the break-in code, and they hand results to each other and change course when something fails. Google says criminals and spy groups are now taking open-source penetration-testing kits, which were built so companies could probe their own defenses, and turning them on real victims. It names ARTEX, CAIRN, PentAGI, HexStrike AI and CyberStrikeAI among the tools in play.

The ways in are old ones: stolen passwords and systems nobody patched. What changes is speed and scale. Software that never sleeps can try the same trick on server after server while its owner is doing something else.

Two monitors with blurred code side by side on a desk
02

The exploit_agent case: a hacking tool that reads patch notes

The case that made Google sound the alarm sat on a staging server the attacker had left open to the internet. On it, Google’s researchers found a multi-agent system called exploit_agent, connected to several commercial and self-hosted AI models, along with data already taken from victims and scripts for the later stages of a break-in.

Its trick is known as patch diffing. When a popular piece of web software gets a security update, the tool lines up the old code against the new code. Whatever the developers changed points almost straight at the weakness they were closing. The AI then drafts an exploit (code that abuses the flaw) and runs it against sites that haven’t installed the update yet. Google says this is the first time it has seen attackers automate both finding and using flaws with no person in the loop.

The tooling was written in Russian. Victim records on the server point to an operator mostly interested in education, media and entertainment websites in France. One comment in the setup file even told users to put their access keys in the file itself and not in the chat window, a small sign of how ordinary this kind of tool has already become.

Two months before this report, Google’s previous quarterly tracker said it had not yet seen fully autonomous attack pipelines aimed at real targets. One quarter later, that line has been crossed.

03

Why patch speed now decides who gets hacked

Attackers have long preferred known flaws, often called n-day bugs, over brand-new ones. Once a fix is public, much of the research is done for them, and lots of systems stay unpatched for weeks. Google argues AI makes this route even more tempting, because the homework is small and focused: two versions of some code, the release notes and whatever public research exists on the bug.

A recent case shows how short the window has become. Two flaws in PaperCut print management servers were chained together and exploited as zero-days in late August. After disclosure, security researchers used AI models to produce working exploits within a day. A campaign spotted four days after disclosure, believed to rely on AI-built exploits, went on to hit hundreds of organizations, and in one break-in reached full control of a company’s network accounts within minutes.

There’s a messy side effect, too. When PHP announced two flaws in what came to be called PHP2Shell, hundreds of supposed proof-of-concept exploits showed up soon after, most of them written by AI. Defenders had to test each one to separate the real ones from the junk, which cost them time they didn’t have.

The FBI has raised its own flag. In an advisory last week, it said several groups have been weaponizing the ARTEX testing framework since August and pointing it at healthcare groups, defense firms, government agencies, tech companies and banks.

Phone, laptop and Wi-Fi router on a home desk during an update
04

What Google expects over the next six months

Google’s researchers put rough dates on the threat. Within about three months, they expect attackers to run unsophisticated attacks that need little or no human help and lean mostly on teams of AI agents. Within about six months, they think sophisticated attacks that run entirely on their own are likely to show up.

The report also lists other ways criminals are putting AI to work:

  • LLMjacking: stealing a company’s AI account keys and cloud access, then reselling cut-price access to powerful models through underground proxy services.
  • Chatbot-powered phishing kits: a Chinese-language phishing service called Jieyou added an AI assistant, so buyers with no technical skill can redesign fake login pages just by typing what they want.
  • Booby-trapped coding assistants: a suspected North Korean group tracked as UNC6836 approached crypto companies through fake recruiter profiles, then altered Claude Code settings inside internal code repositories so the app pulled down a backdoor every time it launched.
  • Influence campaigns: state-backed groups and hired contractors now routinely use AI to mass-produce fake personas and posts across social platforms.

Google says it disabled the accounts and assets tied to misuse of its own Gemini models and used what it learned to sharpen Google DeepMind’s safety classifiers.

There is one piece of good news. Fully automated tools are noisy. Because they act without a careful human steering them, they leave a heavy forensic trail, and for now that makes them easier for defenders to catch.

05

What this means for your phone, laptop and website

For most people, an AI cyberattack won’t look any different from an ordinary hack. There’s no robot on your screen. There’s a hijacked account, a scam email with flawless grammar, or a site you trust suddenly serving something nasty. The defense is mostly the same as before, you just get less time to act. These are standard security habits, not steps taken from Google’s report:

  1. Turn on automatic updates for your phone, computer, browser and apps, and install pending updates the same day instead of snoozing them.
  2. Remember the quiet devices. Your Wi-Fi router, network storage drive and smart home hub run software too. Check their apps or settings pages for updates, and retire gear that no longer gets any.
  3. Use passkeys or two-factor sign-in on email, banking and work accounts. AI-written phishing no longer gives itself away with clumsy spelling.
  4. Be suspicious of job offers that rush you toward code or files, especially in crypto and tech. Fake recruiters opened the door in one campaign Google described.
  5. If you run a website or a small business server, aim to patch web apps, plugins and print or file servers within a day or two of a security release, and take anything off the open internet that doesn’t need to be there.

The rule of thumb is simple: the moment a security fix is published, the clock starts for attackers as well as for you.

06

Quick answers

Are AI cyberattacks already happening?

Yes. Google says both criminal and state-backed groups have used AI agent frameworks in real intrusions, and it found one system that discovered and exploited flaws with no human involved.

Can AI hack my phone on its own?

Not in the way movies suggest. The automated tools Google described go after internet-facing servers with known, unpatched flaws. Keeping your devices updated closes most of those doors.

Is AI being used for defense too?

Yes. Google says it runs its own automated defense systems that watch for threats and respond at machine speed, and it shut down the accounts it caught misusing Gemini.

You're reading the quick version.