A suspected Russian-speaking operator used a multi-agent AI framework to convert security patches into working exploits and steal data from victim organizations without any human involvement, Google Threat Intelligence Group reported on November 16 in its fourth-quarter 2026 AI threat tracker1,14. The FBI warned separately that the ARTEX penetration-testing framework poses a serious threat to organizations' networks, with some intrusions progressing from initial targeting to privileged access in under three hours12,13. Google did not establish that the operation was coordinated by authorities in Moscow6.
Google Threat Intelligence called it the first known case of an AI agent autonomously turning security patches into working exploits15. The group said the attack marks a significant advance in how attackers use large language models for fully autonomous vulnerability discovery and exploitation compared to Q3. Google researchers said adversary tradecraft has transitioned beyond manual prompting to deploying autonomous multi-agent offensive frameworks to execute multi-stage intrusions at greater operational velocity3.
How exploit_agent works
Threat hunters found an exposed staging server hosting the framework, dubbed exploit_agent, which integrates with multiple frontier and local LLMs. The framework used Russian-language tooling and was configured to conduct continuous differential analysis of popular web application software.
exploit_agent compares code before and after security updates, identifies potential vulnerabilities, develops exploits, and executes them without human involvement. One documented target was Spectra (Ultimate Addons for Gutenberg), a WordPress plugin, in version 2.19.257. The server also held data stolen from victim organizations, along with multiple post-compromise scripts that allowed attackers to perform internal reconnaissance, brute-force passwords, install webshells, and remove webshells left by other attackers.
Victim telemetry on the server suggested the operator targeted French websites associated with the media and entertainment industries and the education sector9.
Broader offensive AI landscape
Google's researchers said they observed several threat actors experimenting with autonomous offensive frameworks, including financially motivated actors and suspected espionage threat actors with a PRC nexus. Beyond exploit_agent, attackers have used other AI-powered offensive frameworks including Cairn, CyberStrikeAI, Hexstrike, and PentAGI. One cyberattack campaign targeting organizations in the aerospace sector used the Cairn penetration-testing framework. Suspected North Korean hackers used Claude Code to deliver malware.
The FBI warned in an advisory that multiple threat actors started weaponizing ARTEX since August, launching attacks against organizations in the financial, government, technology, defense, and healthcare sectors. Amazon reached the same conclusion after monitoring its own decoy systems, noting that none of the techniques in these campaigns were new, even as the speed of execution changed4.
Google's report also noted that the underlying attack methods have not changed; what has changed is how quickly attackers can chain them. Google Threat Intelligence said these frameworks currently trade stealth for velocity, generating noisy traffic, recognizable identifiers, and traces in the code they execute10. Google SecOps has rolled out behavioral and volume-based detection rules for ARTEX.
Timeline projections
Google researchers estimated that in the next three months, threat actors will be able to launch unsophisticated attacks that require little to no human interaction and rely mostly on multiple AI agents. Their projection is that sophisticated, fully autonomous agentic attacks are likely to emerge in the next six months.
ANALYSIS The exploit_agent case collapses the window between patch disclosure and active exploitation to machine speed: because the framework automates the comparison of pre- and post-patch code, defenders can no longer assume a published fix buys meaningful remediation time.
Google also observed that hundreds of purported proof-of-concept exploits, described as probably AI-generated, appeared shortly after PHP disclosed two flaws. Separately, Anthropic recently shortened its coordinated vulnerability disclosure window from 90 days to 30 days for high- and critical-severity issues, citing the growing speed at which frontier models discover software vulnerabilities[2].
ANALYSIS The flood of AI-generated proof-of-concept code alongside autonomous exploitation frameworks compounds the defender's problem: security teams must now triage both real and fabricated exploits while the genuine attack timeline shrinks.
Google Threat Intelligence said hackers' use of AI has reached "a critical inflection point".