Ravie Lakshmanan
Nov 16, 2026

In a significant escalation of the digital threat landscape, Google revealed on Monday that the adversarial use of artificial intelligence by malicious actors has officially moved past the phase of manual prompting. According to new insights published by the Google Threat Intelligence Group (GTIG), sophisticated threat groups are no longer relying solely on isolated queries to large language models. Instead, they are deploying fully autonomous, multi-agent offensive frameworks to execute complex, multi-stage intrusions at unprecedented operational speeds.

This technological pivot marks a fundamental shift in how cyber attacks are conceived, deployed, and scaled. Rather than depending exclusively on human-driven inputs for every step of a cyber intrusion, modern threat groups are leveraging agentic frameworks that can plan, execute, adapt, and chain together varied attack paths automatically. While the foundational mechanics of cyber intrusions—such as exploiting network boundaries, harvesting credentials, and lateral movement—remain fundamentally the same, the velocity and scale at which these operations unfold have accelerated dramatically.

Security researchers have identified several prominent open-source offensive frameworks being abused in the wild, including ARTEX, CAIRN, CyberStrikeAI, and PentAGI. These tools are increasingly integrated into attacker workflows to automate discrete phases and specific operational tasks during live intrusions. By adopting these agentic systems, malicious actors can orchestrate attacks at machine speed, drastically compressing the timeline between initial reconnaissance and total network compromise.

The adoption of these technologies is not limited to a single niche of cybercrime. Observations from the threat intelligence community indicate that a diverse array of adversaries is experimenting with these frameworks and operational harnesses. This includes financially motivated cybercriminal syndicates as well as suspected state-sponsored espionage actors aligned with China, who are continuously seeking ways to optimize their operational efficiency and evade traditional perimeter defenses.

Among the specific tools and services cataloged by researchers are specialized platforms like ZeroHollow, an inference service explicitly utilized for exploit development and complex LLM routing, and Agent Zero, an open-source agent framework characterized by its autonomous code execution and dynamic tool generation capabilities. These services provide the computational backbone and architectural flexibility required to sustain multi-agent cyber operations over extended periods without constant human intervention.

Simultaneously, threat actors are weaponizing large language models and agentic pipelines to streamline some of the most labor-intensive aspects of traditional vulnerability research. One notable area of misuse involves automating the reverse engineering of software patches via differential analysis. By feeding source code variations, pre-release preview channels, release notes, security research, and proof-of-concept code into LLM-driven pipelines, adversaries can rapidly weaponize newly disclosed vulnerabilities almost immediately after they become public.

Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches

This capability was vividly demonstrated earlier this September during attacks targeting PaperCut NG/MF flaws, where threat groups utilized hundreds of automated AI interactions to dissect patches and generate functional exploits at scale. By focusing their research on a finite, structured set of source materials, threat actors can effectively collapse the window between patch disclosure and active exploitation, leaving defenders with a dangerously narrow timeframe to secure their environments.

Beyond the execution of direct intrusions, the commercial ecosystem supporting these cyber threats has also evolved to embrace artificial intelligence as a modular force multiplier. A prime example highlighted in the Google report is Jieyou, a Chinese-language phishing platform that has integrated an automated LLM chatbot assistant. This feature is designed to allow customers and low-skilled affiliates to configure complex phishing panels and modify existing templates using simple, natural language commands, removing the traditional technical barrier to entry.

Rather than manually coding payloads, editing HTML templates, or drafting custom social engineering pretexts from scratch, users of the Jieyou platform can now interact with an integrated chatbot to dynamically customize pretext lures and phishing interfaces on the fly. This evolution reflects a broader industry trend where Phishing-as-a-Service providers utilize AI assistants to democratize high-end cybercrime, enabling less technically proficient actors to mount convincing, scalable phishing campaigns with minimal effort.

As threat actors increasingly adopt automated systems, the infrastructure supporting artificial intelligence itself has become a primary target for malicious monetization. Google’s findings also warn of a surge in "LLMjacking" attacks, wherein adversaries actively target enterprise AI infrastructure for illicit profit. In these campaigns, threat groups compromise cloud environments to steal LLM API credentials and consume high-end cloud compute resources.

Once harvested, unauthorized access to these expensive AI models and computing pools is frequently resold through underground proxy services and specialized marketplaces, such as APIMart and JetNight, often referred to within the cybersecurity community as LLM mixers. By reselling discounted access to stolen enterprise AI infrastructure, cybercriminals create a sustainable revenue stream that further funds their broader automated operations.

The convergence of open-source agentic frameworks, automated patch weaponization, and LLM-driven phishing services represents a formidable challenge for enterprise security teams. The widespread availability of these technologies has effectively lowered the barrier to entry, allowing malicious groups to integrate new exploits rapidly, scale their operations autonomously, and execute sophisticated attacks with unprecedented velocity. As the threat landscape continues to adapt to the capabilities of machine-speed attacks, organizations must evolve their defensive strategies to detect and neutralize autonomous agent activity before intruders can establish a foothold within their networks.

By Asro