Google Documents First Fully Autonomous AI Exploitation Framework
Google's GTIG reports the first AI framework that finds and exploits vulnerabilities without humans, plus live agentic attacks in Asia and $1M cloud abuse cases.
By Grace Kim
3 min read

What's News
- Google's November 16 report documents the first LLM framework automating vulnerability discovery and exploitation with no human in the loop.
- Between late August and early October, an actor ran a customized PentAGI platform against organizations across Asia.
- LLMJacking incidents produced over $100,000 in fraudulent usage in a day, reaching $1 million cumulatively in the worst cases.
- Google named PentAGI, CAIRN and CyberStrikeAI among agentic tools experimented with by financially motivated and suspected espionage actors.
- Google has deployed behavioral detection rules and published Mandiant hardening guidance.
Google's Threat Intelligence Group says it has documented the first known case of an AI framework automating vulnerability discovery and exploitation end-to-end, with no human in the loop. The finding anchors a November 16 threat report from GTIG and Mandiant that also confirms live agentic attacks against organizations across Asia.
The report marks a shift in attacker behavior. Google's investigators describe adversaries moving away from one-off AI prompting toward multi-agent systems that can run reconnaissance, vulnerability analysis, exploitation and parts of post-compromise activity on their own.
The findings rest on Google's own investigations and telemetry. They describe observed tools and activity — they do not claim every user of the named open-source frameworks is malicious.
What did Google actually find?
The centerpiece is an exposed staging server running a multi-agent framework Google calls exploit_agent. The system compared software code before and after security patches, identified possible vulnerabilities, drafted exploits and executed them automatically.
The server held more than an exploit engine. Google said it also contained scripts for internal reconnaissance, brute-force activity, installing web shells and removing other attackers' web shells.
Google called this the first instance it has identified in which an LLM framework automated both vulnerability discovery and exploitation without ongoing human direction. That sentence deserves a second read: reconnaissance through exploitation, no operator required.
How active are these attacks in Asia?
More than a lab curiosity. Between late August and early October, GTIG tracked an actor using a customized version of the open-source PentAGI platform for systematic reconnaissance and vulnerability scanning against organizations across Asia.
Google assessed that financially motivated actors and suspected espionage operators have experimented with a wider toolset: PentAGI, CAIRN, CyberStrikeAI and other agentic testing tools.
The report also covers LLMJacking — incidents in which attackers stole credentials or compromised cloud environments to drain organizations' AI and computing resources. The numbers are sharp. Some incidents produced more than $100,000 in fraudulent usage in a single day. In the most severe cases Google investigated, cumulative abuse reached $1 million.
Takeaway: the tooling is open-source, cheap and already in field use against Asian targets.
Who is behind it?
GTIG attributed AI-supported reconnaissance, stolen-data analysis and influence operations to several government-linked or suspected state-backed actors. Google flagged the limits itself: those attribution claims remain the company's assessments and are not independently established by the report's public evidence alone.
So the state-actor angle is directional, not proven. The autonomous exploitation framework and the Asian scanning campaign are the concrete, observed pieces.
Can defenders catch this today?
Yes, for now — and that's the practical message of the report. Google said autonomous offensive frameworks currently trade stealth for speed, producing traffic and code artifacts that defenders can detect.
Google has responded on two fronts: it has deployed behavioral detection rules and published hardening guidance through Mandiant.
That's the opening defenders have. Agentic tools are fast but noisy — detection rules tuned to their artifacts can catch them today. The open question is how long the noise advantage lasts as the frameworks mature; Google's own timeline shows the gap between experimentation and real-world attacks narrowing from months to weeks.
Source: technode.global
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,206 articles