Google's threat researchers say they have watched an attack framework do something defenders have long feared: read a security patch, work out what it fixed, write an exploit for the flaw, and run that exploit with no person at the keyboard. In its AI Threat Tracker published on November 16, the Google Threat Intelligence Group, working with Mandiant, called the moment a critical inflection point and said it was the first time it had identified threat actors using large language model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.
The evidence came from an exposed staging server hosting a multi-agent system Google dubbed exploit_agent. The framework connects to multiple frontier and local large language models and runs continuous differential analysis of popular web application software. It compares code before and after a patch announcement, analyzes the difference to identify the vulnerability the patch closed, drafts an exploit, and automatically executes it. Alongside data taken from targeted victims, researchers found scripts for what comes after a break-in, including internal reconnaissance, brute-forcing, installing webshells, and removing other attackers' webshells.
Victim telemetry on the server suggested the operator was specifically interested in French websites in the media and entertainment industries and in the education sector. Google named no group and no country for the operator. The only origin clue it published was in a configuration file, where Russian-language comments told the user to enter LLM API keys for auto-exploitation in the file itself rather than in a chat window. Google also gave no success rate for the framework and no count of sites it compromised. One example it printed showed the system inspecting access-control code in Spectra, the Ultimate Addons for Gutenberg WordPress plugin, at version 2.19.25.
The finding does not stand alone. Google said threat actors are also abusing open-source multi-agent penetration-testing frameworks in live intrusions, naming ARTEX, CAIRN, CyberStrikeAI, and PentAGI, and that it has seen experimentation with HexStrike AI. It tracked a customized platform built on PentAGI conducting systematic reconnaissance and vulnerability scanning against organizations across Asia from late August to early October 2026, and it observed CAIRN used against targets including the aerospace sector in the United States. As supporting evidence for automated exploitation pipelines, the report cited late August attacks on PaperCut print management servers, where two vulnerabilities, CVE-2026-81578 and CVE-2026-82078, were chained for unauthenticated remote code execution and, according to GreyNoise, a campaign days after disclosure likely used LLMs to build exploits and target hundreds of organizations.
The wider numbers explain why patch timing now matters so much. Google said overall vulnerability disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose from a 2025 average of 8 a month to 11 a month in 2026 before spiking to 22 in August. By chaining patch analysis, exploit drafting, and execution together, the researchers wrote, threat actors can dramatically reduce the window between patch disclosure and weaponized exploitation, putting pressure on the familiar enterprise habit of patching in days or weeks.
Google was also clear about the current limits. Today's agentic frameworks trade stealth for speed, generating heavy traffic and distinct code-level traces that raise the chance of detection. Espionage groups that prize persistence may avoid such noisy tools for now, while financially motivated criminals may accept lower success rates in exchange for far greater reach. For enterprise teams, the practical lesson is narrower and urgent: a published patch is now also a map for machines, so internet-facing systems need fixes, and detection, measured in hours rather than in the next maintenance cycle.