Google has flagged what it describes as the first observed cyberattack in which an artificial intelligence system carried out vulnerability discovery and exploitation without any human intervention.

The finding comes from a new report by the Google Threat Intelligence Group (GTIG) and Mandiant, titled "Agentic Threat Activity Crosses a Significant Threshold", published on November 16. Google said it marks a critical inflection point in how cybercriminals use AI.

Until now, attackers used AI to help with individual tasks. The report says they are now deploying autonomous systems that can plan and execute multiple stages of an attack on their own, compressing intrusion timelines that once took days or weeks into minutes.

"For the first time," the report said, researchers observed threat actors "using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop."

The exploit_agent framework

At the centre of the report is a multi-agent framework Google dubbed exploit_agent, which its researchers found running on an exposed staging server left open to the internet.

The framework connects to multiple frontier and locally hosted large language models. It continuously compares the code of popular web applications before and after each security patch announcement, works out what flaw the fix reveals, drafts an exploit and runs it automatically.

The server also held data stolen from victim organisations, along with scripts designed to support activity after gaining access, including internal network reconnaissance, brute-force attempts and installing webshells to maintain access to compromised systems.

Based on the server's activity, Google said the operator appeared to be interested in French websites associated with the media, entertainment and education sectors. Russian-language comments in the operation's code pointed to a Russian-speaking operator, according to CyberScoop's reporting on the findings.

A wider shift to agentic attack tools

The report is not about one tool. Google said adversaries are now deploying open-source, multi-agent penetration testing frameworks, among them ARTEX, CAIRN, CyberStrikeAI and PentAGI, and that it has seen experimentation with HexStrike AI.

ARTEX, an autonomous penetration testing framework pulled from GitHub in October, has been tied to automated attacks on banks in Southeast Asia. The FBI warned in an alert on November 13 that criminals have used it since at least August 2026.

Google tracked one actor using the CAIRN framework against the US aerospace sector, and another running a customised PentAGI platform to scan organisations across Asia for vulnerabilities between late August and early October.

In one cited case, attackers exploiting newly disclosed vulnerabilities moved from an initial foothold to full domain administrator control within minutes.

Patch-to-exploit windows are shrinking

Google said overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 11 a month in 2026, up from eight a month in 2025, before jumping to 22 in August.

The report points to two flaws in PaperCut print management servers disclosed in late August 2026. Security researchers built working exploits using large language models within a day of the announcement, and within four days a campaign believed to have used AI-generated exploits was detected targeting organisations.

The danger is that autonomous systems can exploit publicly disclosed vulnerabilities, known as n-days, before companies finish applying security updates. "Attackers don't need zero-days," said John Hultquist, chief analyst at the Google Threat Intelligence Group, because AI lets them turn a patch into an attack faster than defenders can deploy the fix.

Google said it expects less human involvement and better stealth as the frameworks mature, and forecast that simple multi-agent attacks with almost no human involvement could emerge within three months, with sophisticated fully autonomous attacks likely within six months.

What defenders can do

Google said its response includes model-level safeguards and security tools aimed at vulnerability discovery and hardening against machine-speed attacks, including its Google AI Threat Defense offering built on Gemini, Wiz, CodeMender and Mandiant.

The company said it disabled the accounts and assets tied to misuse of its Gemini API that it found during the investigation, and that DeepMind used the findings to strengthen its classifiers.

For organisations, the report's central message is that patching speed now matters more than ever: the gap between a fix being published and an attack built from it is collapsing from weeks to hours.