Google Threat Intelligence Group and Mandiant have documented the first known instance of threat actors using a fully autonomous AI framework to discover and exploit software vulnerabilities without human intervention. The finding, published November 16 in a report titled Agentic Threat Activity Crosses a Significant Threshold, marks a paradigm shift toward intrusions occurring at machine speed.
The Exploit_agent Framework
Researchers discovered the multi-agent framework, dubbed exploit_agent, on an exposed staging server. The system connects to frontier and local large language models to compare code before and after patch announcements, identify flaws, draft exploits, and execute them automatically. The server also contained exfiltrated victim data, reconnaissance scripts, brute-force tools, and webshell deployment and removal utilities.
Victim telemetry indicated the operator targeted French media, entertainment, and education websites. Code comments within the framework pointed to Russian speakers, though Google has not attributed the activity to a specific group or nation-state.
Differential Analysis and the Shrinking Patch-to-Attack Window
The exploit_agent framework leverages differential analysis of patches — comparing unpatched and patched code to pinpoint the vulnerability — to weaponize n-day flaws rapidly. GTIG notes that attackers find patched vulnerabilities easier to exploit with LLMs than zero-days because they can focus on a finite set of source materials such as code versions and release notes.

Overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits grew only marginally from 2025, averaging 11 per month against eight per month the previous year, though they spiked to 22 in August 2026. The PaperCut print management server vulnerabilities illustrated the trend: researchers generated working exploits for two chained flaws within a day of disclosure, and GreyNoise observed a threat actor likely using LLMs to build and test exploits just four days post-disclosure before hitting hundreds of organizations, expanding one foothold to domain administrator within minutes.
Multi-Agent Offensive Frameworks Proliferate
Adversaries are increasingly deploying autonomous multi-agent offensive frameworks including ARTEX, CAIRN, CyberStrikeAI, and PentAGI to execute multi-stage intrusions at greater operational velocity. ARTEX has emerged as the most used framework, primarily by Chinese-speaking groups. GTIG observed CAIRN deployed against the US aerospace sector and a customized PentAGI platform conducting systematic reconnaissance and vulnerability scanning across organizations in Asia. A Chinese group was also observed building its own automated attack framework, likely for espionage operations.
Additional tools in the ecosystem include ZeroHollow, an inference service for exploit development and LLM routing, and Agent Zero, an open-source agent framework featuring autonomous code execution and dynamic tool generation.
Broader Threat Landscape Trends
State-sponsored groups adopting LLMs include multiple Chinese APTs such as Leaky Castle, Iran-linked APT42 (also known as Calanque Ion), and North Korean hackers supporting IT worker activity and cryptocurrency attacks such as Midnight Neptune (UNC1069). Phishing-as-a-Service operators are also integrating LLMs; the Jieyou phishing service, active since March 2026, uses a model-agnostic chatbot allowing clients to configure phishing pages.
Financially motivated actors continue to profit from LLMJacking — stealing LLM credentials and cloud resources from compromised enterprise environments and reselling access through underground proxy services such as LLM Mixers.
Defensive Posture
Jonathan Ong, senior analyst at Omdia, characterized the development as an acceleration of existing attack techniques rather than novel methods. Keith Prabhu, CEO of Confidis, noted limited evidence that AI makes individual attacks more successful, but warned that attackers can launch more attempts in less time, potentially increasing successful compromises. Ong added that the frameworks typically generate considerable detectable activity, meaning enterprises with properly configured and monitored EDR, XDR, or MDR capabilities should still identify much of the malicious activity.
Google says it is hardening models against misuse, disrupting bad actor projects and accounts, and operationalizing security at machine speed through its autonomous Google AI Threat Defense architecture and BeyondZero.
Sources
- Google Spots Fully Autonomous AI Attacks For the First Time
- GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold | Google Cloud Blog
- AI Framework Autonomously Weaponized Flaws in Real Attacks, Google Says
- Google spots the first hacker AI system that finds and exploits flaws autonomously
- Fully Automated AI Cyberattack 2026: What Google Just Found – Memeburn
Photo: Chris Ried / Unsplash (Unsplash License)





