KEY POINTS
- Google's GTIG says a suspected Russian-speaking operator used an AI agent to turn patches into exploits with no human in the loop, a first in its tracking.
- An exposed staging server hosted a framework called exploit_agent, along with stolen victim data and post-compromise scripts.
- Google expects more threat actors to adopt autonomous offensive frameworks because of their low barrier to entry.
A suspected Russian-speaking operator used an AI agent to convert security patches into working exploits without any human involvement, according to the Google Threat Intelligence Group (GTIG). The Register reported that Google describes this as the first such case it has seen, and that the operator went on to steal data from victim organizations.
Google documented the activity in its fourth-quarter 2026 AI threat tracker. The company said the case marks a significant step forward from the third quarter in how attackers use large language models for fully autonomous discovery and exploitation of vulnerabilities.
Investigators found an exposed staging server that hosted a multi-agent framework for vulnerability research and weaponization. The framework, named exploitagent, works with several frontier and local models and relied on Russian-language tooling. It was set up to carry out continuous differential analysis of what the report calls popular web application software.
In practice, that means the system compares code before and after a security update, identifies possible vulnerabilities, develops exploits and then runs them automatically. No person had to step in at any stage, according to the report.
The server also held data taken from victim organizations. In addition, it stored multiple post-compromise scripts that could be used for internal reconnaissance, brute-forcing, installing webshells and removing webshells left by other attackers. Victim telemetry on the server suggested the operator focused on French websites linked to the media and entertainment industries, as well as the education sector.
The report describes other cases in which attackers moved past manual prompting of LLMs and instead deployed autonomous, multi-agent offensive frameworks for multi-stage intrusions. These involved both financially motivated criminals and cyber-espionage groups. Besides recent attacks involving the open source penetration-testing tool ARTEX, attackers have used frameworks including CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. In one instance, GTIG saw PentAGI used together with proxy services to target the aerospace sector.
Google's researchers said several threat actors have separately experimented with these frameworks, including financially motivated groups and suspected espionage actors with a link to China. Citing the low barrier to entry, the ease of adding new exploits and the potential for massive scaling, they wrote that it is highly likely such actors will keep using these tools in campaigns against organizations worldwide.