Google Reports AI Used in Vulnerability Exploitation
November 19, 2026 · 2 min read
Also published in Español, Italiano, Nederlands, Português (Brasil), Svenska
Google reported on November 16 that it has confirmed a case in which an AI system is used by actual attackers to automatically discover software vulnerabilities and execute attack code without human intervention. This finding indicates a potential reduction in the time between the public disclosure of vulnerability patches and actual attacks, thereby necessitating quicker responses from businesses.
The discovery was made by the Google Threat Intelligence Group (GTIG) and Mandiant in their latest AI threat report. Previous investigations had identified some automation in vulnerability scanning and attack procedures, but this marks the first confirmation of a system that operates entirely without human involvement in the exploitation process.
GTIG identified an "exploit_agent" mechanism installed on externally accessible servers. This system collaborates with multiple large language models (LLMs) to continuously compare pre- and post-patch code of widely used web applications. It analyzes potential vulnerabilities and automates the creation and execution of attack code.
Additionally, the same server contained scripts used for internal investigations of breaches, as well as for establishing backdoors for attackers to regain access. GTIG suspects that the operator was targeting websites in the media, entertainment, and education sectors in France.
According to Google, when targeting known vulnerabilities with published patches, attackers can narrow down their focus using pre- and post-patch code and update information. This suggests that using AI could enable more efficient creation of attack code compared to searching for unknown vulnerabilities from scratch.
While the speed of attacks increases, the automation also leaves traces such as communication records and distinctive code, which can aid in detection of the attacks. Google noted an increase in AI-generated code claimed to be useful for attacks, even if it does not function as intended, adding a burden on defenders to verify its authenticity.
Moreover, Google has implemented measures to disable related accounts and assets involved in abuse of the Gemini AI. The insights from this investigation have also been reflected in improvements to AI models and detection capabilities for misuse.