Models

Google says fully automated cyberattacks have arrived

Google’s Threat Intelligence Group says attackers have fully automated vulnerability discovery and exploitation. The “exploit_agent” framework points to French media, entertainment and education sites.

Google warns that fully automated cyberattacks have arrived and will get worse fast
Google warns that fully automated cyberattacks have arrived and will get worse fastFirmbee / pixabay
By Elena Vasquez5 min read

Why it matters

  • GTIG says it observed fully automated vulnerability discovery and exploitation for the first time, with no human in the loop.
  • The exposed exploit_agent framework connects to multiple language models, according to the GTIG AI Threat Tracker.
  • GTIG says the system compares code before and after patches, identifies vulnerabilities, develops exploits and runs them automatically.
  • Russian-language configuration comments suggest a Russian-speaking operator; server data points to French websites in media, entertainment and education.
  • GTIG says the attacks currently trade stealth for speed and leave forensic traces that make them easier to detect.

Google’s Threat Intelligence Group says it has observed attackers fully automating vulnerability discovery and exploitation “for the first time, with no human in the loop.”

The finding centers on an exposed server. It runs an AI framework named “exploit_agent.” According to GTIG’s AI Threat Tracker, the framework connects to multiple language models. It carries the process from code analysis to exploit execution.

That matters because GTIG describes a complete attack chain inside one system. The research stakes are practical. The group says human intervention no longer sits between finding a flaw and using it.

The evidence offers a direct look at that chain. It also comes with clear limits. The available account identifies no operator, individual victim or named language model. No date appears for the observed activity. GTIG presents the finding as a first observation without attaching a date.

What does exploit_agent do?

GTIG says the system compares software code before and after patches. In the group’s workflow, that comparison reveals changes introduced by a fix, identifies vulnerabilities, develops exploits and runs them without any human analyst reviewing each stage.

The reported workflow has four steps.

  • Compare software code before and after patches.
  • Identify vulnerabilities in the comparison.
  • Develop exploits for those vulnerabilities.
  • Run the exploits automatically.

Patch comparison gives the process a specific starting point. GTIG says the framework looks on both sides of a change. Then it turns that view into vulnerability detection and executable code.

The crucial distinction, in GTIG’s account, is not merely that attackers used AI. The group says the human was outside the loop. The system generated and executed exploit code on its own.

What did the operator leave behind?

GTIG found Russian-language comments in the server’s configuration files. The group says this suggests the operator speaks Russian. The language clue does not identify a person, organization or government.

The same investigation produced server data about the apparent target set. According to GTIG, that data points to French websites in three sectors. Those sectors are media, entertainment and education.

The evidence goes beyond the framework’s code. GTIG says the server data points to French sites. That gives the finding a targeting dimension. The report does not say whether those sites were compromised or only selected as targets.

Those categories are broad. The available account does not list individual sites or companies. The evidence indicates a French and sector-specific focus. It leaves the size and timing of the operation unspecified.

The pattern still matters for research and defense. GTIG ties language clues, configuration artifacts and server data into one account. Those traces connect the tooling with an apparent intent, yet they stop short of identifying the operator behind the Russian-language comments and French target data.

Why is this a critical turning point?

GTIG calls the shift “a critical turning point.” The phrase refers to the first observed case of attackers fully automating discovery and exploitation. GTIG says no human remained in the loop.

The concrete change is the joining of stages. The framework compares patched code, identifies weaknesses and develops exploits. It then executes them inside the same automated workflow.

For defenders, the relevant interval runs from finding a flaw to running an exploit. GTIG’s account says that sequence can proceed without human approval.

Google’s broader warning is blunt: fully automated cyberattacks have arrived and will get worse fast. That forecast rests on the observed system. GTIG’s judgment is that the shift will spread.

The defensive stakes follow from the system’s design: under the workflow GTIG describes, no manual review sits between comparing patched code and automatically running an exploit. The source does not claim every organization faces equal risk. It shows the capability exists, making the warning more than a theoretical scenario.

Can automated attacks still be detected?

GTIG says the current attacks sacrifice stealth for speed. The group reports that forensic traces make the systems easier to detect. That tradeoff gives defenders something concrete to examine.

Speed appears in the automated chain from patch comparison to exploit execution. Stealth, in the observed implementation, moves at a different pace. The system leaves evidence behind.

GTIG’s evidence includes an exposed server, configuration comments and server data. Those artifacts gave the group material to trace the operation’s apparent focus, from Russian-language configuration comments to server data pointing toward three French target sectors.

The defensive opening is not invisibility. GTIG says the automation is currently detectable. Investigators and security teams have traces to examine. The underlying workflow can still become more autonomous.

The source does not say how long that tradeoff will last. Google warns that automated attacks will get worse fast, according to the report’s framing. Speed is already obvious. Stealth remains imperfect.

What remains unknown?

Much of the significance depends on details the available account does not provide. It does not identify the language models, operator, targeted websites or attack count.

The summary gives no timeline for the French activity. It does not explain how long the exposed server operated. Those omissions limit conclusions about the operation’s scale and duration.

The Russian-language comments support only a limited inference. GTIG says they suggest a Russian-speaking operator. They do not establish a Russian state connection or a particular criminal group.

The French target data is similarly suggestive rather than exhaustive. GTIG uses “suggests” for the target assessment. That wording preserves uncertainty about whether recorded intentions match every action taken.

That uncertainty matters because Google’s warning covers a trend. The concrete evidence described here comes from one exposed server and one framework. The finding establishes a capability and a pattern. It cannot quantify every use of that capability.

Because the evidence comes from one exposed system, it may reveal more about the operator’s process than about the wider threat, and GTIG provides no figures for measuring that spread. Without counts or timelines, readers cannot judge prevalence from this finding alone.

What happens next?

GTIG’s warning is not that automation has already erased the attacker’s footprint. Its current finding is the opposite. Speed has arrived before stealth. Forensic traces remain visible.

Future operators may try to preserve automated speed while reducing forensic exposure. Until GTIG reports otherwise, the exposed server, configuration comments and target data remain the clearest signs that the loop can now close without a human.

Original: cloud.google.com

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

708 articles

Related articles

  1. Google Threat Intel: AI Agents Now Turn Security Patches Into Exploits
  2. Google finds fully autonomous exploit pipelines in the wild
  3. Google Reports First Autonomous AI Attack That Finds and Exploits Flaws
  4. Google Cloud ships new security tools for autonomous AI agents
  5. AI found a Google zero-day for $4.24 in 13 minutes in 2026

« Previous articleNext article »