Google Threat Intelligence Group (GTIG) says it has observed threat actors using multi-agent AI systems to discover and exploit software vulnerabilities without a human in the loop, marking the first time the company has seen the full process automated in live attacks. The findings appear in Google and Mandiant’s November 16 report, which tracks the growing use of agentic frameworks across cybercrime and espionage.

One case involved an exposed staging server running a custom framework called exploit_agent. The system connected to multiple commercial and locally hosted language models, continuously compared source code from popular web applications before and after software updates, and used the differences to identify potential flaws. It then drafted exploit code and executed it automatically. Google said the server also contained data exfiltrated from victims, internal reconnaissance tools, web shells and brute-force scripts. Telemetry suggested the operator focused on French websites in media, entertainment and education.

Google’s report also shows the framework analyzing an access-control check in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25. The example demonstrates code analysis and does not establish that users of the plugin were compromised.

Google said threat actors are using open-source multi-agent tools such as ARTEX, CAIRN, CyberStrikeAI and PentAGI. GTIG has observed CAIRN and PentAGI deployed in live activity: CAIRN was used in an intrusion involving U.S. aerospace targets, while a customized PentAGI-based platform conducted reconnaissance and vulnerability scanning against organizations in Asia between late August and early October. Google has also seen experimentation with CyberStrikeAI and HexStrike AI. The report’s findings follow an earlier WorkOS incident in which an ARTEX-powered agent escaped a testing sandbox and accessed an internal application.

READ
Check Point Gateways Hit in Attacks That Wiped ESXi Files

These systems can discover and exploit low-complexity flaws, generate single-use payloads and adjust their activity based on execution feedback. Google said some attackers are also using AI to reverse-engineer software patches, compare code changes and weaponize newly disclosed vulnerabilities, shortening the time between a fix becoming public and an exploit being deployed.

Google’s report also details LLMjacking, where attackers steal cloud credentials and compute resources to run or resell access to AI models, a practice covered in Abijita’s report on AI cloud-billing abuse. Google said the automated frameworks currently generate recognizable forensic traffic and code-level traces that can make them easier to detect, but it expects the tools to become more stealthy as their capabilities improve. The report also notes that AI-generated proof-of-concept exploits often fail to work, adding verification work for security teams after vulnerability disclosures.


Buy ExpressVPN with PayPal or Credit Card
Advertisement