Skip to content
Google reports first AI

Google reports first AI

Biz.Chosun • November 17, 2026

Hackers have begun using artificial intelligence (AI) agents to analyze security patches and turn them into working exploit code to launch large-scale automated attacks. Analysts say hacking is speeding up as AI finds vulnerabilities and writes exploit code on its own without human intervention.

Google's Threat Intelligence Group (GTIG) and Mandiant released an "AI Threat Tracking Report" on the 17th, saying attackers are fully automating the reverse-engineering process of comparing code changes before and after security patches with large language models (LLMs) to pinpoint the fixed vulnerabilities.

Although a security patch (the fixed vulnerability) had been released, attackers weaponized so-called N-day vulnerabilities—flaws not yet applied to systems—to carry out mass automated attacks on internet-exposed servers, the report said. In an intrusion targeting the U.S. aerospace industry, the threat actors used a self-directed multi-agent penetration tool called CAIRN.

"Hackers' tradecraft is evolving from manually entering prompts to using autonomous multi-agent attack frameworks such as Artex, CAIRN, CyberStrikeAI, and PentAGI," GTIG said. "That lets them move through multi-stage intrusions faster." With AI agents handling system intrusion and exploitation, human involvement is decreasing and the attack tempo is accelerating, it said.

In this report, GTIG offers an in-depth look at how attackers across cybercrime, espionage, and information operations (IO) are using Generative AI to shorten timelines, scale up campaigns, and execute multi-stage intrusions at machine speed.

GTIG said it also confirmed for the first time an exploitation case with a fully human-out-of-the-loop workflow. An LLM framework called "exploit_agent," built around Russian-language tools, continuously compared code before and after patches to find vulnerabilities and automatically developed exploit code without human input.

In influence operations, investigators found a case where attackers used the Google Gemini application programming interface (API) to build a platform that manages more than 100,000 fake accounts. The platform ran cloud phone farms—a cloud-based system that operates large numbers of virtual smartphones—to manage phone-linked accounts. It even had features that mimic smartphone camera-specific defects and distortions so generated content would look real, it was found.

"In information operations, state-backed threat groups, Influence-as-a-Service contractors, and fake account networks that monetize political content now routinely use LLMs to mass-produce synthetic media and to create localized personas that automate distribution across multiple platforms," the report said.

In targeted spying, China-linked actors were found to analyze and process stolen data with LLMs to produce intelligence reports. North Korea-linked actors, for their part, reportedly used AI-generated personas to target cryptocurrency companies. Some of those personas bore Google's AI watermark SynthID, according to GTIG. Attempts to tamper with settings for autonomous coding assistants were also observed.

Extracted Entities