Google Warns: Soon Cyber Attacks Will Be Managed by Artificial Intelligence, Not Humans
· Tech · Cem Koyluoglu
Cyber attackers now use AI-based attack infrastructures to automate the detection of security vulnerabilities and the exploitation of those vulnerabilities.
TL;DR Cyber attackers are now deploying AI‑driven, agent‑based attack workflows that can autonomously discover and exploit vulnerabilities at machine speed, targeting a wide range of sectors. Major frameworks such as ARTEX, CAIRN, and Claude Code are being weaponized, and the trend is expected to grow into fully autonomous attacks within months.
Key Highlights • AI‑powered frameworks (ARTEX, CAIRN, CyberStrikeAI, HexStrike AI) enable fully automated vulnerability discovery and exploitation. • High‑profile attacks include CAIRN on U.S. aviation and aerospace, and Claude Code malware distributed by suspected North Korean actors. • The FBI warned of ARTEX usage across finance, public, technology, defense, and health sectors. • Google’s AI Threat Tracker identified a multi‑agent system, “exploit_agent,” used to target French media, entertainment, and education sites. • State‑backed groups and commercial suppliers are increasingly employing AI for influence operations and “LLMjacking.”
Artificial‑Intelligence‑Powered Attack Workflows
Cyber adversaries are now employing agent‑based workflows that run on AI‑driven attack infrastructures, allowing them to detect and exploit security weaknesses at machine speed. While AI was initially used experimentally in attack chains, Google’s Threat Intelligence Group (GTIG) has observed that both financially motivated and state‑backed threat actors are adopting AI on a large scale.
High‑Profile Examples
In a recent assault on the United States’ aviation and aerospace sector, attackers used CAIRN, an open‑source, multi‑agent penetration‑testing framework. Separately, hackers suspected of North Korean affiliation distributed malware through the Claude Code platform. Researchers anticipate that within the next three months, threat actors may launch attacks that require minimal or no human interaction, relying largely on multiple AI agents and avoiding complex orchestration. In the following six months, the emergence of “complex and fully autonomous” agent attacks is expected.
FBI Alert on ARTEX Framework
The FBI issued a warning last week that several threat actors have been weaponizing the ARTEX AI penetration‑testing framework since August, targeting organizations across finance, public, technology, defense, health, and other sectors.
Using AI to Find and Exploit Vulnerabilities
Google reports that attackers are systematically abusing open‑source penetration‑testing packages—including ARTEX, CAIRN, CyberStrikeAI, and HexStrike AI—to conduct fully automated attacks. These frameworks enable adaptable assaults against low‑complexity vulnerabilities that are susceptible to large‑language‑model (LLM) exploitation and exposed internet endpoints.
In GTIG’s AI Threat Tracker, hackers employed a multi‑agent system called “exploit_agent” to automate the entire process of discovering and exploiting vulnerabilities against French media, entertainment, and education websites. Google stated:
“GTIG first detected that threat actors were using LLM frameworks to fully automate vulnerability discovery and exploitation without human intervention in the last quarter.”
“Specifically, GTIG identified an externally exposed preparation server hosting the multi‑agent vulnerability research and weaponization framework named exploit_agent integrated with multiple leading and local LLMs.”
The infrastructure was located on an attacker‑owned internet‑facing preparation server, which also contained stolen victim data, reconnaissance information, web‑shell tools, and brute‑force attack scripts. The system analyzes software updates in popular web applications to identify vulnerability indicators and automatically attempts exploitation.
Google warns that the use of such frameworks drastically shortens the window defenders have to patch systems and can generate misleading indicators. Researchers highlighted that shortly after PHP announced two vulnerabilities, hundreds of AI‑generated proof‑of‑concept exploitation codes appeared in the recent PHP2Shell example. Defenders were forced to verify these codes and develop countermeasures, placing a heavy burden on security teams.
Other AI‑Related Trends
Beyond automated attacks, Google identified multiple assaults by UNC6836—suspected of being linked to “Midnight Neptune”—that targeted cryptocurrency companies in North America and Europe using a modified version of Claude Code. After the initial breach, attackers altered the Claude Code settings to trigger the download of a SOMBERMEME backdoor.
Google reports that state‑backed groups, commercial suppliers, and political networks now routinely employ AI in influence operations. Researchers note that “LLMjacking,” the unauthorized use of large‑language‑model resources, has become a widespread issue. Attackers increasingly exploit this new revenue stream, imposing significant costs on affected organizations.