A recent report issued by the Google Threat Intelligence Group (GTIG) revealed a new and dangerous phase of cyberattacks, where it observed for the first time attackers using AI-based frameworks capable of automating the entire attack lifecycle, from vulnerability discovery to exploitation, without the need for direct human intervention.
During monitoring operations, the group identified a framework called exploit_agent, which was discovered on an exposed experimental server. This framework was designed to compare web application code between patched versions and unpatched versions, then identify vulnerabilities, develop exploits, and execute them automatically. The discovered server also contained stolen data from victims and scripts for post-exploitation activities, such as reconnaissance and deploying webshells, with indications of targeting the education, media, and entertainment sectors in France.
Offensive Frameworks Operating at "Machine Speed"
The report explained that attackers, including state-sponsored groups, have begun adopting multi-agent frameworks to increase the speed and scope of their operations. Among the most prominent of these frameworks are:
- Artex: Considered the most widely used framework, its use is primarily associated with Chinese-speaking groups.
- Cairn: Its use was observed in attacks targeting the aerospace sector in the United States.
- PentAGI: Used in systematic reconnaissance and vulnerability scanning across organizations in Asia.
- CyberStrikeAI and HexStrike AI: Other frameworks being tested and developed by attackers.
Additionally, the use of tools such as ZeroHollow for inference services and exploit development was noted, as well as the open-source framework Agent Zero, which is characterized by its ability to execute code and generate tools dynamically and independently.
Expansion of the Scope of AI-Powered Threats
It is not limited to direct hacking; the use of Large Language Models (LLMs) has extended to include other areas:
- Phishing-as-a-Service: Such as the Jieyou service, which allows customers to customize phishing models via chatbots.
- Influence-as-a-Service (IaaS): The use of fully automated platforms to manage disinformation campaigns, from creating fake personas to interactive posting across platforms.
- AI Resource Theft (LLMJacking): Operations targeting the theft of access credentials to AI models to exploit their computing capabilities, which caused financial losses amounting to one million dollars in one case.
Why Does This Development Represent a Real Danger?
The fundamental danger lies in the shift toward what experts have described as "machine-speed attacks." Previously, vulnerability discovery and exploitation required specialized human teams and long periods of time; now, however, AI drastically reduces the required cost and expertise, placing advanced capabilities within the reach of small and unprofessional groups.
This development puts organizations, especially critical infrastructure such as water plants and hospitals, in a losing race if they rely on traditional patching cycles that occur monthly. An automated attacker can exploit a vulnerability as soon as it appears, while a human defender needs days or weeks to respond. Therefore, the report emphasizes the need for a rapid transition toward automated defenses that can detect and respond at the same speed as automated attacks.