---
title: Google reports artificial intelligence framework automatically converted security patches into exploits
url: https://www.dataloco.com/en/google-reports-artificial-intelligence-framework-automatically-converted-security-patches-into-exploits
published: 2026-11-19T21:21:24+00:00
language: en
section: Security
source: https://www.techtalkthai.com/google-says-an-ai-framework-turned-security-patches-into-exploits-with-no-human-in-the-loop/
organizations: Google, Google Threat Intelligence Group, Huntress Labs, GreyNoise Intelligence
publisher: Dataloco
---

# Google reports artificial intelligence framework automatically converted security patches into exploits

Google Threat Intelligence Group disclosed that attackers have for the first time deployed an artificial intelligence framework operating without human involvement to transform security patches into exploits for attacks. The framework, named exploit_agent, was discovered on an unprotected test server connected to both frontier and locally hosted large language models.

According to the report, exploit_agent continuously compared code in popular web applications before and after each patch announcement. Upon identifying potential vulnerabilities, the system automatically wrote attack commands and executed them. Stolen data from victims was stored on the same server alongside scripts for brute force attacks, web shell installation, and removal of web shells left by other attacker groups. Telemetry from victims indicated the operators showed particular interest in French websites related to media, entertainment, and education. A comment in the framework's configuration file, written in Russian, instructed the system administrator to enter a key "here, not in chat."

Two months earlier, Google Threat Intelligence Group reported that fully autonomous attack processes had not yet been deployed against real targets. That situation changed in the recent quarter, according to a new report titled "Agentic Threat Activity Crosses a Significant Threshold." Because sources for already-patched vulnerabilities are limited, the group believes attackers may find using large language models to target n-day vulnerabilities more effective than pursuing zero-day vulnerabilities.

The speed of closing the gap between disclosure and functional exploit creation was demonstrated by two vulnerabilities in PaperCut Software's print management server. Security researchers and Huntress Labs used a large language model to create a functional exploit within one day of disclosure. A campaign tracked by GreyNoise Intelligence four days after disclosure reportedly relied on large language models to create exploits before attacking hundreds of organizations, with one attack point reaching administrator access within minutes.

Open source agent-based penetration testing frameworks are being used to automate various stages of ongoing intrusions. During one intrusion targeting the United States space sector, attackers deployed a multi-agent framework named CAIRN. Another group ran a customized platform built on PentAGI between late August and early October to scan for vulnerabilities in organizations across Asia.

Clear forensic traces from unrestricted automated operations currently make these frameworks easier to detect. According to Google Threat Intelligence Group, current influence operations use models for more than just writing posts, including phone farms on cloud systems. A ready-made platform that a commercial contractor is building with assistance from Gemini API was designed to control more than 100,000 social media accounts. Prompts in another platform focused on political attitudes across Europe and the Middle East were identified as instructing the creation of intentional camera defects to make generated images resemble smartphone photographs.

Among state-sponsored attackers, profile images used by fake recruiters from a group tracked as UNC6836 and suspected to be North Korean, deployed to approach cryptocurrency organizations in North America and Europe, were found to carry Google's SynthID watermark. The group later modified a .claude/settings.json file in an internal code repository to trick each use of Anthropic's Claude Code into downloading their SOMBERMEME backdoor. Google stated it closed accounts and assets associated with the detected misuse of Gemini. Google DeepMind used the findings to strengthen its classifiers.
