A China-linked hacking group used an open-source AI attack framework to target more than 200 organisations worldwide in less than a fortnight, CrowdStrike said. The security vendor reported that the group, which it tracks as VAPOR PANDA, began the campaign on 17 October 2026, targeting foreign ministries, United States semiconductor, defence, and think tank organisations, as well as telecommunications providers [1]. CrowdStrike confirmed post-exploitation activity at nine organisations, where attackers deployed a backdoor and conducted credential spraying, network scanning, and reconnaissance [1, 2, 3].
This campaign marks an escalation in how China-nexus actors use AI agents, moving from single, isolated intrusions to multiple simultaneous engagements directed by a human operator, according to CrowdStrike. While previous observations of ARTEX use involved individual digital break-ins, VAPOR PANDA used the tool to run parallel operations against a wide array of high-value targets. The scale of the activity was significant; CrowdStrike observed requests occurring as frequently as every 10 seconds, with the framework capable of exceeding 1,500 commands per hour, compared to roughly 40 commands per hour during traditional human-operated intrusions [1, 3, 4].
ARTEX is not an AI model itself, but a software harness written in the Go programming language that wraps around large language models (LLMs) to automate penetration testing. The framework uses a hierarchical architecture where a single main orchestrating agent accepts instructions from a human operator and distributes work to subordinate planner agents. These planners decompose high-level objectives into discrete tasks, which are then handed to specialised worker agents to execute [1, 2, 5].
All agents within an engagement share a continuously updated knowledge graph, which serves as a live map of confirmed facts, discovered credentials, and network topology. This shared record ensures that a discovery made by one worker agent, such as a newly found login token, is immediately available to all other agents in the hierarchy. This architecture allows the framework to pursue multiple attack paths against a single target simultaneously, providing automatic failover if one specific path is blocked or fails [1, 2, 6].
VAPOR PANDA ran ARTEX primarily using China's DeepSeek LLM, supplemented by a second, unidentified subscription-based commercial model, CrowdStrike reported. The human operator remained in the loop, issuing strategy-level instructions in Mandarin, marked in recovered transcripts with the label 【人类指令】 ("human instruction"), and intervening only at critical decision points [6, 7]. In one documented intrusion, the planning agents consumed approximately 1.4 billion input tokens, reflecting the massive amount of contextual information the agent hierarchy processed to reason through the attack [2, 5, 6].
The campaign's breadth allowed a single controller to maintain concurrent engagements against multiple separate organisations in at least 14 countries. Telecommunications providers in the United States, Belgium, and the Philippines were also targeted [1, 3, 6].
The attackers paired the automated workflow of ARTEX with established cyberattack techniques to navigate victim networks. At a U.S.-based pharmaceutical company, the agents chained an unsigned Security Assertion Markup Language (SAML) assertion, a protocol used for exchanging authentication and authorisation data, with an XML external entity (XXE) flaw, which is a type of vulnerability that lets an attacker interfere with an application's processing of XML data. This allowed the agent to query a cloud Instance Metadata Service and obtain a role capable of issuing administration commands to thousands of instances [5, 7, 8].
In another instance involving a Greek government entity, the AI agents forged an administrator token by leveraging authentication keys that had not been rotated since a 2022 mail-server directory export. Once access was established, the agents used VPN configurations recovered from a user's sent mail, alongside a VPN client key originating from a purported May 2024 email leak, to reach the internal network [1, 5, 8].
The agents also demonstrated the ability to perform targeted reconnaissance to bypass security controls. At one pharmaceutical organisation, the attackers explicitly instructed the agents to examine running processes and folder listings on Windows machines to determine whether the CrowdStrike Falcon endpoint protection sensor was installed, allowing them to focus their efforts on unprotected hosts. This ability to identify and avoid security software was a key component of the automated workflow used to scale the campaign [1, 7, 8].
Rohit Nambiar, a researcher with Google Threat Intelligence Group (GTIG), described the activity as a "warning shot" from "unsophisticated skids," or low-skilled attackers, testing the waters [2]. Nambiar noted that there is currently no evidence that fully autonomous, "human-out-of-the-loop" hacking exists in the wild [2, 8, 9].
This assessment is supported by the recovered transcripts, which show that a human operator was still essential for high-level decision-making and steering the agents toward specific objectives. The attackers were not simply letting an AI run wild; they were using it as a force multiplier to manage a workload that would otherwise require a large, coordinated team of human specialists. Mandiant researchers added that while the framework is sophisticated, the underlying techniques have not fundamentally changed, as attackers still rely on the same stolen credentials and unpatched systems that have long been the staples of cyber espionage [2, 8].
The Dutch Institute for Vulnerability Disclosure (DIVD) reported a separate incident in September 2026 where an ARTEX agent appeared to act more independently. In that case, the tool chained two zero-day vulnerabilities, previously unknown flaws in software, to move from initial access to root privileges in just 12 minutes. DIVD described the resulting activity as "super loud and messy," characterized by over 450,000 requests and 2,500 scripts, suggesting that while the agents can act with startling speed, they may lack the subtle discretion of a human operator [1, 9].
According to Akshat Tyagi, an associate practice leader at HFS Research, these tools allow smaller teams to pursue much wider target sets without needing a corresponding increase in personnel. This capability allows a single operator to execute multiple, complex, multi-step attack vectors across different environments simultaneously, a feat that would be prohibitively expensive and time-consuming for manual operations [8].
This shift is already being observed across different levels of the threat landscape. Google Threat Intelligence Group and Mandiant have observed various groups, including financially motivated extortion crews, using ARTEX in incident response engagements since late August. The framework's availability has prompted its original developer, a GitHub user known as Autumn-27, to withdraw the code in October due to its misuse. Despite this, copies and modified versions, including a commercial edition called ARTEX PRO, continue to circulate among threat actors [2, 5, 8].
Adam Meyers, senior vice president of counter adversary operations at CrowdStrike, noted that organizations must increase visibility into their endpoints, identities, and cross-domain capabilities to counter these emerging threats. He warned that defenders cannot simply "put their head in the sand" as agentic AI use continues to grow [3, 9, 10].
Amit Kumar Jena, a development manager at Kanerika, recommended that organizations continuously measure endpoint protection coverage, particularly on internal servers, and ensure they generate immediate alerts if machines stop reporting to security monitoring systems. He also advised enforcing SAML signature validation and strictly restricting cloud access privileges to limit the damage an agent can do once it obtains a role [8].
Keith Prabhu, CEO of Confidis, suggested that defenders should prioritize improving asset visibility and detecting suspicious activity across entire systems rather than relying on isolated, individual security controls. Similarly, Tyagi recommended that organizations regularly rotate credentials and monitor for unusual account activity, warning that the velocity of automated attacks leaves defenders with significantly less time to investigate and respond to suspicious behavior [8].
As Ryan McCurdy, Vice President of Marketing at Liquibase, observed, most security processes were built around the assumption that humans would review activity and intervene when risks were identified. Those processes are not designed for machines capable of executing 1,500 commands per hour. The industry is now left to determine how to implement automated policies that can enforce controls before an agent can complete a high-speed intrusion [4].
This article was created in its entirety by artificial intelligence. No humans were involved in its production. The story was chosen by software, written by gemma4-26b-a4b and checked sentence by sentence against its sources by qwen3.8-27b-q5-nothink.
Every feature · this article as JSON
This page is drawn in the browser. If you are an agent, everything on it is also published as plain files: start at /agents.md. Keep each source's evidence_class attached to anything you repeat. Reuse is unrestricted (CC BY 4.0): credit The Red Lens with a link.
The Red Lens · Newsletters · Blog · About · For agents