AI Framework Autonomously Weaponized Flaws in Real Attacks, Google Says
State-sponsored and cybercrime groups are increasingly adopting generative AI, including multi-agent offensive frameworks, to increase the speed and scale of their operations, Google Threat Intelligence Group (GTIG) reports.
For the first time, GTIG has identified threat actors using an AI framework to automate both vulnerability discovery and exploitation without human intervention.
The multi-agent framework, called exploit_agent, was discovered on an exposed staging server. It was designed to compare patched and unpatched web application code, identify vulnerabilities, develop exploits, and execute them automatically.
The server also contained data stolen from victims and scripts for post-compromise activities, including reconnaissance and webshell deployment. According to GTIG, evidence pointed to an interest in French media and entertainment websites, as well as the education sector.
Attackers are using frameworks such as Artex, Cairn, CyberStrikeAI, and PentAGI to increase the velocity of their multi-stage intrusions, and rely on LLMs and agentic frameworks to reverse engineer patches at machine speed and rapidly weaponize n-day vulnerabilities.
“Threat actor adoption of these frameworks represents a paradigm shift towards intrusions occurring at machine speed, with the automated chaining of complex attack paths scaling operations significantly,” GTIG notes in its latest AI threat tracker report.
Artex has emerged as the most used AI framework for threat actors, mainly Chinese-speaking groups, but Google and Mandiant also observed attackers using and experimenting with open source frameworks such as Cairn, PentAGI, CyberStrikeAI, and HexStrike AI.
GTIG observed Cairn being used in activity targeting the aerospace sector in the US, and PentAGI being used for systematic reconnaissance and vulnerability scanning across organizations in Asia.
Not only have threat actors experimented with these frameworks and harnesses, but GTIG also observed a Chinese group building its own automated attack framework, likely for use in espionage operations.
“These observations additionally extend to the use of other tools and services, such as ZeroHollow, an inference service used for exploit development and LLM routing, and Agent Zero, an open source agent framework which features autonomous code execution and dynamic tool generation,” GTIG notes.
According to Google, threat actors will continue to adopt these tools and their intrusions will become nearly fully agentic.
“As the sophistication of these tools and underlying LLMs increases, we anticipate not only reduced human involvement in these operations but also increased stealth and sophistication of the activity,” GTIG notes.
State-sponsored hackers expand AI use
State-sponsored groups known to have adopted 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 running attacks against cryptocurrency organizations, such as Midnight Neptune (UNC1069).
Phishing-as-a-Service (PhaaS) operators too are using LLMs. The Jieyou phishing service, active since March 2026, has a model-agnostic chatbot that allows clients to configure phishing panels and modify templates.
According to Google, threat actors such as state-sponsored groups, commercial Influence-as-a-Service (IaaS) vendors, and inauthentic political monetization (IPM) networks are adopting LLMs for content generation and infrastructure design, which improves the quality of long-running influence campaigns.
“This broad adoption has measurably improved the output quality of some long-running influence campaigns while enabling IPM networks to scale political clickbait that resembles the thematic focus, quality, and volume historically characteristic of state-backed IO campaigns. However, it is important to note that improvements in content quality do not necessarily equate to greater impact or success,” GTIG notes.
In August 2026, Google discovered an end-to-end IaaS platform that automated the entire IO campaign lifecycle, from persona creation to interactive posting across platforms. A separate commercial contractor was developing an AI-powered platform designed to manage over 100,000 accounts across multiple platforms.
Other threat actors are hunting for credentials that provide access to LLMs to monetize them as part of LLMJacking operations. This type of activity can generate significant usage and billing increases, with cumulative resource abuse reaching $1 million in one case.
The surge in LLMJacking has led to the proliferation of LLM Mixers, illicit marketplaces selling access to LLM services and underlying proxy infrastructure and routing traffic through pools of stolen credentials and compute instances. They offer access to multiple models, including frontier and jailbroken ones.
According to GTIG, the wide adoption of LLMs in malicious operations introduces forensic noise that degrades OPSEC and stealth. LLMs can reduce operational reliability, and stealth operators are likely to selectively employ noisy tools for speed and scale while looking to maintain long-term persistence.
“Conversely, we assess with high confidence that actors less concerned with operational stealth, including many cybercrime actors, will continue incorporating autonomous offensive tools into their operations, given that the increased reach of their campaigns will likely compensate for any decreases in success rates,” GTIG notes.
“For US critical infrastructure, the danger isn’t just more attacks, but attacks at machine speed. Water utilities, hospitals, and regional electric utilities that struggle to patch on a monthly cycle can’t win that race manually. AI is dramatically lowering the cost and expertise required to conduct sophisticated cyberattacks. Tasks that once required well-resourced teams, from discovering vulnerabilities to exploiting them and moving through networks, can increasingly be handed off to agents. That puts powerful capabilities within reach of smaller, less sophisticated groups,” said Merlin chief strategy officer Matt Hartman.
“This report is an early warning of where things are headed. We should assume these capabilities will spread. Nation-state adversaries and criminal operators are watching closely to see what works. Defenders must do two things at once: nail the cybersecurity fundamentals today and accelerate the shift toward automated defenses that can detect and respond at machine speed,” Hartman added.
Originally published on SecurityWeek