Google warns AI could be used to launch automated cyberattacks without human involvement
A report from the Google Threat Intelligence Group (GTIG) notes that hackers are now employing autonomous AI agents and large language models (LLMs) to automate cyberattacks without human involvement.
Google analysts warn that autonomous agentic attacks could become common within the next three to six months.
Autonomous Attack Frameworks
Reports indicate that financially motivated cybercriminals as well as suspected China-linked espionage groups are testing open-source AI agent tools. These include ARTEX, CAIRN, CyberStrikeAI, PentAGI, and Agent Zero. These tools enable the coordination of multiple stages of an intrusion since they rely on an automated decision loop. This allows them to execute code dynamically at high speeds.
Attackers are using LLMs to constantly scan for software updates and compare patched versus unpatched code. Then, by reverse-engineering these fixes, they are able to exploit vulnerabilities before security teams have the time to apply patches. Google observed this pattern with PaperCut NG/MF software and in attacks against French media and education websites. The AI models analyzed public updates and developed working exploit code.
LLMjacking
Running powerful AI models can require substantial computing resources and involve significant costs. To avoid these costs, hackers are breaking into enterprise systems where they steal AI credentials and gain unauthorized access to computing resources. This involves hackers scanning the internet for accidentally exposed API keys or stolen account credentials belonging to companies that pay for this access.
They resell this stolen access through underground proxy services called LLM mixers. These platforms are used to sell thousands of stolen accounts and AI access designed to conceal users’ identities to cybercriminals.
AI-Powered Phishing
The report notes that AI is making online scams much easier to orchestrate. Therefore, low-skilled criminals without any special expertise can easily trick high-value targets. Phishing-as-a-service (PhaaS) platforms like Jieyou now include built-in chatbots. A scammer with little or no coding experience can simply ask the AI to rewrite the website’s HTML/CSS code.
Moreover, North Korean state hackers were found using AI-generated profile photos that Google identified as AI-generated through its SynthID watermarking technology. They used these pictures to pose as legitimate recruiters or crypto investors. This made it easier for them to trick software developers into downloading fake project files that hijack Claude Code. This instantly installs a malware backdoor called SOMBERMEME onto the developer’s device.
