Attackers have for the first time deployed artificial intelligence frameworks that autonomously convert security patches into working exploits with no human involved, Google's Threat Intelligence Group revealed Monday — meaning the same class of machine intelligence that reads your email and tracks your searches is now writing attack code on its own.
The threshold crossed here isn't subtle. Two months ago, Google said it had never observed fully autonomous attack pipelines in the wild. That's over. The machines are now operating at machine speed, and the window between a patch announcement and a working exploit has collapsed from weeks to hours.
The framework, dubbed exploit_agent, was discovered on an exposed staging server wired into multiple large language models, according to SiliconANGLE. Its function: continuously compare application code before and after each security patch, identify the flaw, and automatically draft and execute an exploit. Stolen victim data sat on the same server alongside brute-force scripts and tools for installing and removing rival attackers' webshells. The operator targeted French media, entertainment, and education websites. Configuration files were written in Russian, with one comment telling the operator to enter keys "right here, not in the chat."
The speed is what matters. When PaperCut Software disclosed flaws in its print management servers, security researchers at Huntress Labs used large language models to generate working exploits within a day of disclosure. A campaign tracked by GreyNoise Intelligence just four days later — believed to have relied on LLM-built exploits — hit hundreds of organizations. One foothold reached domain administrator within minutes.
Google's report, titled "Agentic Threat Activity Crosses a Significant Threshold," notes that because the source material for a patched flaw is finite, attackers may find it more efficient to point LLMs at known vulnerabilities than to hunt for zero-days. The Hacker News reported that open-source frameworks like CAIRN, PentAGI, and CyberStrikeAI are lowering the barrier to entry, letting low-skilled actors chain attacks at scale. During one intrusion against the U.S. aerospace sector, a threat actor used the multi-agent framework CAIRN. Another actor ran PentAGI to scan Asian organizations for vulnerabilities between late August and early October.
The infrastructure isn't just being used for code. Google found influence operations using cloud phone farms and a turnkey platform built with the Gemini API designed to run more than 100,000 social media accounts. Prompts for generated photos included deliberate camera flaws so images would pass as smartphone shots, targeting political sentiment across Europe and the Middle East. A suspected North Korean group tracked as UNC6836 used fake recruiter profiles carrying Google's own SynthID watermarks to approach cryptocurrency organizations in North America and Europe.
The Hacker News also reported that threat actors are stealing LLM credentials from compromised enterprise environments and reselling access through underground proxy services called LLM mixers — services like APIMart and JetNight. A Chinese-language phishing platform called Jieyou has integrated an LLM chatbot so low-skilled users can configure phishing campaigns using natural language commands without technical expertise.
For now, Google says the noisy forensic trail left by unrestricted autonomous execution makes these frameworks easier to detect. That's cold comfort. Detection is not prevention, and the machines are only going to get quieter.
Google built the AI. Google's API powers the influence farms. Google's watermarks show up on North Korean phishing profiles. The question isn't whether autonomous attack AI exists — Google just confirmed it does. The question is who, if anyone, can control what they've built.








