The Finding
Google's Threat Intelligence Group (GTIG) published findings on November 16, 2026 documenting a qualitative shift in offensive cyber operations: threat actors are now deploying autonomous AI frameworks that identify, weaponize, and exploit software vulnerabilities with minimal human involvement. This is not a theoretical capability — GTIG observed these frameworks operating in real attacks against production targets.
The Frameworks
GTIG identified four active autonomous offensive AI frameworks operating in threat actor infrastructure: ARTEX, CAIRN, CyberStrikeAI, and PentAGI. These frameworks perform LLM-driven patch differential analysis — automatically comparing patched and unpatched versions of software to identify and codify the vulnerability — and then generate working exploits. GTIG observed this process applied to PaperCut NG/MF vulnerabilities in real attacks, with weaponization occurring within hours of patch publication rather than the days or weeks traditionally required.
Supporting Infrastructure
Two tools support the offensive AI ecosystem. ZeroHollow is an inference service that routes exploit development tasks across multiple LLM backends, optimizing for speed and cost. Agent Zero is an open-source autonomous agent with native code execution capabilities, used for running generated exploits against targets. Additionally, GTIG documented a Chinese-language phishing-as-a-service platform called Jieyou that uses an LLM chatbot to configure phishing campaigns via natural language commands, dramatically lowering the technical barrier for phishing operations.
The Strategic Shift
The traditional model of defensive security assumed that organizations had days between a patch release and active exploitation to assess, test, and deploy. That window has collapsed. When AI frameworks can analyze a patch and generate a working exploit in hours, patch management SLAs measured in days become meaningless for critical vulnerabilities. The assumption of human-in-the-loop on the attacker side — which created predictable delays — is now obsolete for adversaries operating these frameworks.
What This Means for Your Security Program
Critical vulnerability patching must move to a 24-hour SLA for CVSS 9.0+ vulnerabilities affecting internet-exposed or laterally-reachable systems. Detection engineering must assume compromise occurs faster than patching. Automated vulnerability scanning and patch deployment pipelines are no longer nice-to-have. And the signal that an exploit exists should now be assumed at the moment of patch publication for any high-profile vulnerability.