Google says attackers used AI to find and exploit flaws with no human in the loop for the first time
On November 16, 2026, Google's Threat Intelligence Group reported that in the past quarter it found attackers using LLM frameworks to find and exploit software flaws with no human in the loop, which it calls a first. Its main example is exploit_agent, found on an exposed staging server, which compared code before and after patches, drafted exploits and ran them. Google names no actor behind exploit_agent, and the only targets it describes for the tool are French media and entertainment websites and the education sector.
What Google reported
Google’s Threat Intelligence Group published its AI Threat Tracker on November 16, 2026. It says that in the past quarter it identified threat actors using LLM frameworks “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
The clearest example is a tool Google calls exploit_agent, which it found on an exposed staging server. The server held a multi-agent framework that connects to several frontier and local language models. According to the report, the tool compares a web application’s code before and after a patch is announced, drafts an exploit for the flaw the patch fixed, and runs it without human intervention. The server also held scripts for internal reconnaissance, password brute-forcing and installing webshells, along with data stolen from victims. One script removed other attackers’ webshells.
Google says victim telemetry on the server suggests the operator was interested in French websites tied to media and entertainment, and in the education sector. It gives no dates for those attacks, no victim names and no count.
A configuration file on the server, where the operator entered model API keys, carried Russian-language comments. One translates as “LLM API keys for auto-exploitation (exploit_agent pilot)”. CoinDesk’s headline led with the Russian text. Google’s post attributes the tool to no group or country and refers to whoever runs it only as “the operator”.
Other cases in the report
Google says it observed CAIRN, an open-source multi-agent penetration-testing framework, used in intrusions that included US aerospace targets.
The report also describes a suspected Midnight Neptune cluster, in its section on North Korean activity. In intrusions against cryptocurrency organisations in North America and Europe, the cluster edited Claude Code’s settings files so that the tool downloaded a backdoor called SOMBERMEME each time a user started it. Google says SynthID watermarks on the actor’s profile photos confirm that AI was used to build fake professional identities.
On two PHP vulnerabilities, Google says a proof-of-concept video and early patched code appeared before the official patches on October 21, 2026. Hundreds of non-working “exploits” followed, likely generated with LLMs, Google says.
What Google did
Google says the activity set off Gemini’s safety responses and that it then disabled the assets tied to the campaigns. It also says Google DeepMind used what it learned to strengthen its classifiers and models. The post does not say whether any victims were told or whether the exploit_agent server was taken down.
The FBI alert
On November 13, 2026 the FBI issued FLASH-20261113-001, “Threat Overview of ARTEX Autonomous Penetration Testing Framework”. It describes ARTEX as an open-source platform, written in Go, that links to outside language models and lets attackers automate several stages of an intrusion.
The FBI says its developer, who goes by Autumn-27 on GitHub, released it in July 2026 and removed it on October 8. Forks and a commercial version called ARTEX PRO still circulate. Since at least August, the FBI says, “a variety of malicious cyber actors” have used it against financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal organisations.
Citing researchers, the FBI says more than 1,600 ARTEX consoles were publicly exposed as of early November. Citing Microsoft, it says some intrusions went from first targeting to privileged access in under three hours. It also cites vendor reports for attacks on South Korean financial institutions that exposed personal data on about 67,000 people. The alert names no state or group as responsible.
Google’s post lists ARTEX among the frameworks threat actors use and says it has built detection rules for it. It does not say whether exploit_agent is related.
Where the news coverage differs
BleepingComputer’s report on the Google findings, dated November 16, goes further than the Google post we read. It says Google expects fully automated attacks needing little human input within three months and sophisticated autonomous attacks within six. It also uses the labels UNC6836 and PHP2Shell. We did not find the forecasts or either label in Google’s post. Instead of a timeline, Google’s post says it is “highly likely” attackers will keep using these frameworks and that operations will move toward “fully agentic intrusions”. BleepingComputer’s article links to no Google source.
Open
The first The “no human in the loop” claim comes from Google and has not been confirmed by anyone else.
Attribution Google names no actor behind exploit_agent. The Russian-language comments hint at where the operator comes from but do not establish it.
Victims French media and entertainment sites and the education sector are the only targets described for the tool.
The tool Which models exploit_agent called is not stated.
The forecasts The three-month and six-month predictions appear only in BleepingComputer’s account.
The date We have used November 16, 2026, the day Google published, because the attacks themselves are undated.
Sources
-
P
Google Cloud (Google Threat Intelligence Group): GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold
retrieved 2026.11.19 -
P
FBI: FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework
retrieved 2026.11.19 -
S
BleepingComputer: Google warns that fully-automated AI attacks are coming soon
retrieved 2026.11.19 -
S
CoinDesk: Google finds self-running AI hacking tool with Russian comments in its files
retrieved 2026.11.19 -
S
The Record: FBI flash alert says Artex tool automating cyberattacks at 'machine speed'
retrieved 2026.11.19 -
T
ITdaily: Google sees AI agent turn patches into exploits without human intervention
retrieved 2026.11.19
Search used: site:cloud.google.com GTIG AI Threat Tracker agentic threat activity; site:ic3.gov FLASH ARTEX
Updated 2026.11.19. Tier and status definitions are in the methodology.