Google’s Threat Intelligence Group has identified a significant evolution in cyber warfare, confirming that threat actors are now deploying multi-agent artificial intelligence frameworks to conduct autonomous, machine-speed intrusions. For the first time, researchers have observed a framework, dubbed exploit_agent, capable of independently discovering vulnerabilities, developing exploits, and executing attacks without human intervention. This tool was discovered on a compromised staging server that also contained stolen victim data and scripts designed for reconnaissance and the deployment of webshells.
The investigation into the staging server revealed that the exploit_agent framework functions by comparing patched and unpatched versions of web application code to identify security gaps. Once a vulnerability is found, the system automates the creation and execution of an exploit. Google reports that this specific activity targeted the education sector and media and entertainment organizations in France.
Beyond this specific instance, Google has documented a broader trend of attackers adopting various AI-driven offensive frameworks to accelerate multi-stage intrusions. These tools, which include Artex, Cairn, CyberStrikeAI, PentAGI, and HexStrike AI, allow adversaries to reverse-engineer patches and weaponize n-day vulnerabilities at a pace that exceeds manual defense capabilities. Artex has emerged as the most frequently utilized framework, particularly among Chinese-speaking threat groups. Other frameworks have been observed in specific regional campaigns, such as Cairn being used against the U.S. aerospace sector and PentAGI facilitating systematic reconnaissance across Asian organizations.
The shift toward agentic operations is further supported by the use of specialized infrastructure and auxiliary tools. Google identified the use of ZeroHollow, an inference service designed for exploit development and LLM routing, as well as Agent Zero, an open-source framework that enables autonomous code execution and dynamic tool generation. Evidence suggests that some state-sponsored actors are moving beyond off-the-shelf tools, with at least one Chinese group observed developing a custom automated attack framework specifically for espionage.
State-sponsored entities are increasingly integrating these technologies into their established operations. Notable groups identified by Google include the Chinese APT Leaky Castle, the Iran-linked APT42, and the North Korean group Midnight Neptune, also known as UNC1069. These groups are leveraging LLMs not only for technical exploitation but also for content generation and infrastructure design to bolster influence campaigns.
The commercialization of these capabilities has also expanded. Phishing-as-a-Service providers, such as the Jieyou service active since March 2026, are now using model-agnostic chatbots to help clients configure phishing panels and modify templates. Additionally, Google discovered an end-to-end platform capable of automating the entire lifecycle of an influence operation, from creating fake personas to posting content across multiple social platforms. Another contractor was found developing a system intended to manage over 100,000 accounts simultaneously.
A parallel trend involves the rise of LLMJacking, where attackers compromise credentials to gain unauthorized access to LLM services. This activity has led to the creation of illicit marketplaces known as LLM Mixers, which sell access to stolen compute instances and proxy infrastructure. These marketplaces provide access to a variety of models, including those that have been jailbroken to bypass safety filters. In one documented case, resource abuse and billing spikes related to LLMJacking reached $1 million.
Google assesses that while the use of AI can introduce forensic noise that may hinder operational security for some attackers, the speed and scale provided by these tools are likely to outweigh the risks for cybercrime groups. The company anticipates that future intrusions will become increasingly agentic, with reduced human involvement and higher levels of stealth. Security experts emphasize that this transition to machine-speed attacks necessitates a fundamental shift in defensive strategy, moving away from manual patching cycles toward automated, high-speed detection and response systems to protect critical infrastructure.





