Google’s threat researchers say attackers have, for the first time, used an AI framework to turn security patches into working exploits and run them with no human in the loop. The finding, published Monday in the Google Threat Intelligence Group’s (GTIG) fourth-quarter AI Threat Tracker, marks the moment automated hacking stopped being a forecast and became an observed attack.
November 16, 2026
What happened
GTIG found an exposed staging server hosting a multi-agent framework its operator called exploit_agent. The framework, built with Russian-language tooling and wired into several frontier and locally hosted language models, continuously compares the code of popular web application software before and after each security patch announcement. When it spots a likely vulnerability in the difference, it drafts an exploit and executes it automatically — no person reviews, approves, or launches the attack.
The same server held data stolen from victim organizations, alongside post-compromise scripts for internal reconnaissance, brute-forcing passwords, and installing webshells — including scripts that removed webshells left by rival attackers. Victim telemetry suggests the operator was most interested in French websites in the media, entertainment, and education sectors. Google has not attributed the operation to a named group.
Why this is a first
Two months ago, in the previous edition of the same quarterly tracker, GTIG said it had not observed fully autonomous attack pipelines deployed against real targets. That line has now been crossed. The new report, titled “Agentic Threat Activity Crosses a Significant Threshold,” says several separate actors are experimenting with similar frameworks — financially motivated criminals as well as suspected espionage actors with a China nexus — using tools including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI, the last of these observed in activity touching the U.S. aerospace sector.
GTIG’s reasoning is uncomfortable for defenders: because the code changed by a patch is a finite, public source, pointing a model at freshly patched (“n-day”) flaws may be more efficient for attackers than hunting unknown zero-days. The report points to August’s PaperCut print-server flaws, where working exploits followed disclosure within about a day with LLM help.
Why it matters
For anyone running a website or business software, the practical change is timing. The gap between “a patch exists” and “someone is exploiting the flaw it fixes” used to be measured in days of human work; an agent that reads the patch itself compresses that gap toward zero. Patching promptly — the advice security teams have given for years — is no longer just best practice. It is now a race against software that starts the moment the patch ships. That is the “everything AI, tested” reality of late 2026: the same agent technology being sold to defenders is already running, unattended, on the other side.
Frequently asked questions
What is exploit_agent?
It is the name of a multi-agent framework Google found on an exposed attacker server. It compares software code before and after security patches, identifies likely vulnerabilities, writes exploits, and runs them automatically.
Did Google name who was behind it?
No. Google did not attribute the operation to a specific group. The tooling was in Russian, and victim telemetry pointed to French media, entertainment, and education websites.
What should website owners do differently?
Apply security patches as soon as they are available. GTIG’s finding means the window between a patch’s release and automated exploitation attempts is shrinking fast.
Sources: The Register, SiliconANGLE, Google Threat Intelligence Group

