Google confirms AI-driven vulnerability attacks without human intervention
November 19, 2026 · 2 min read
Google Cloud reported on November 16 that its cybersecurity research team, Google Threat Intelligence Group (GTIG), has observed a significant expansion in the realm of attacks utilizing artificial intelligence. For the first time, attackers were seen using AI agents to create exploits from security patches without any human involvement.
The attacks have transitioned to autonomous workflows and multi-agent frameworks. The methods employed by attackers have gone beyond manual prompt inputs, utilizing autonomous multi-agent attack frameworks such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI to execute multi-stage intrusions at a faster operational pace.
GTIG confirmed this quarter the first instance of vulnerability exploitation entirely devoid of human involvement. It was reported that the framework "exploit_agent," which collaborates with multiple large language models (LLMs), continuously compares code before and after patch releases to identify vulnerabilities and automatically develop exploits without human intervention.
Furthermore, it has been revealed that attackers are utilizing open-source AI agents for their intrusion activities. GTIG found that both financially motivated threat actors and those involved in cyber espionage are using open-source agent-based penetration testing tools, such as CAIRN and PentAGI, to automate their attack procedures. Instances were reported where threat actors deployed CAIRN in intrusion activities targeting the aerospace sector in the United States.
Additionally, AI is expanding the scale of information warfare campaigns. In information operations (IO), it has been discovered that LLMs are not only used for content creation but also by state-sponsored groups and commercial "Influence-as-a-Service" vendors for building automated pipelines, designing infrastructure, and avoiding detection.
There have also been cases of platforms being developed that exploit the Gemini API to manage over 100,000 synthetic accounts through cloud phone farms.
Moreover, state-sponsored threat actors are employing AI for espionage and target selection. Attackers linked to China have been observed using LLMs to process stolen data into intelligence products. In activities associated with North Korea, there have been movements to target cryptocurrency-related companies using AI-generated personas, as well as methods of manipulating autonomous coding assistants to implant backdoors.