An attacker’s AI framework was using security fixes as material for working attacks, according to Google Threat Intelligence Group. The suspected Russian-speaking operator’s system compared software versions, developed exploits and executed them without human involvement—a sequence that moves AI beyond answering hacking questions and into carrying out the work.
GTIG documented the discovery in its fourth-quarter 2026 AI threat tracker, The Register reported on November 16. Google described the case as a first such occurrence and an advance over the autonomous vulnerability discovery and exploitation it observed in the third quarter.
Published fixes, automated attacks
Threat hunters found the framework, called exploit_agent, on an exposed staging server. It combined multiple AI agents for vulnerability research and exploit development, connected to both frontier and local models, and used Russian-language tooling. The discovery gave investigators a view of the software supporting the operation, alongside other material stored on the server.
Its configuration called for continuous comparisons of popular web application software before and after security updates. That process, known as differential analysis, identifies changes that may reveal a vulnerability. The framework was set up to turn those findings into exploits and then execute them automatically, rather than stop at an analysis for a person to act on.
The approach focuses on flaws for which a fix already exists. In its explanation of patch-based exploit development, Google contrasted that bounded research task with a less directed search: attackers can work from specific code changes and accompanying disclosures. The Hacker News quoted Google’s reasoning:
Threat actors may find it more efficient to use LLMs to rapidly weaponize patched n-day vulnerabilities, as they can focus research on a finite set of source materials: comparing source code versions (including from pre-release or preview channels), and examining release notes, security research, and POC code related to specific flaws
Google, quoted by The Hacker News
Stolen data alongside the framework
The server held more than research tooling. Investigators also found data stolen from victim organizations and multiple scripts for activity after a compromise. Those files placed the framework alongside evidence of data theft and tools for operating inside compromised systems.
- Internal reconnaissance scripts supported examining a victim’s environment after access had been gained.
- Brute-forcing scripts supported repeated attempts to break through access controls.
- Other scripts could install webshells—tools for remote access through a web server—and remove webshells belonging to other attackers.
Victim telemetry suggested a particular targeting interest, rather than an indiscriminate list of sectors. GTIG said the operator appeared interested in French websites associated with media and entertainment, as well as education-sector websites. That assessment describes apparent targeting, not a confirmed identity for the operator.
From prompting to running intrusion tasks
Google’s wider findings describe attackers moving beyond manual chatbot prompts to frameworks that automate stages of live intrusions. It named ARTEX, CAIRN, CyberStrikeAI and PentAGI among the tools being used. The distinction is operational: the systems help carry out discrete attack tasks and connect stages of an intrusion, rather than only supply text or code on request.
Those experiments span different motives. Google observed financially motivated actors and suspected China-aligned espionage actors using offensive AI frameworks; that wider observation is separate from the suspected Russian-speaking exploit_agent operator. GTIG expects continued use globally, citing low barriers to entry, easy integration of new exploits and the potential to scale operations.
Reader comments
Newest comments first. Replies stay oldest first.