AI framework helps China-linked hackers probe 200 organisations in days

CrowdStrike says the Vapor Panda group used AI agents to coordinate intrusions across at least 14 countries, with confirmed post-compromise activity at nine organisations.

Chinese state-backed hackers have been using an artificial intelligence-enabled penetration testing framework called ARTEX to widen the reach of their operations, according to CrowdStrike, which said the tool has allowed a small team to push multiple intrusions forward at once rather than relying on the slower, manual methods that typically define espionage campaigns. The company attributed the activity to a China-based intrusion group it tracks as Vapor Panda, also known as VexatiousTemper, and said the cluster overlaps with Earth Estries, a threat set previously described by Trend Micro as targeting government-linked organisations and technology companies across several countries.

CrowdStrike said the campaign began on 17 October 2026 and, within days, had touched more than 200 organisations in at least 14 countries. Among the targets were foreign ministries in Latin America, Europe, Asia and the Middle East, as well as US organisations in the semiconductor, defence, pharmaceutical, nuclear and financial sectors. Telecommunications providers in the United States, Belgium and the Philippines were also listed among those affected, underscoring the breadth of the operation and the variety of sectors the actors were willing to probe.

According to CrowdStrike, the attackers used ARTEX alongside DeepSeek and another unidentified commercial large language model to coordinate a planner-and-worker model that could move from reconnaissance to exploitation and lateral movement. Human operators remained involved, but mainly to guide the agents, set objectives and steer the campaign when needed. In one intrusion, the company said, planning agents consumed around 1.4 billion input tokens, a figure that suggests a highly automated effort operating at unusual scale. Post-compromise activity was confirmed at nine organisations, where the attackers deployed BLOODALCHEMY, a ShadowPad derivative, before moving on to credential spraying, network scanning and further reconnaissance.

The recovered transcripts indicate that the operators paired the AI framework with tried-and-tested intrusion techniques. CrowdStrike said the agents attempted to chain an unsigned SAML assertion with an XML external entity flaw to query cloud metadata services, then sought role access that could issue administrative commands across thousands of instances. In other cases, the attackers deployed web shells, exploited an unauthenticated orchestration service to obtain root access in a production container environment, extracted Local Administrator Password Solution credentials and used resource-based constrained delegation to impersonate privileged accounts. The company also said the actors relied on old authentication keys to forge administrator tokens and reach mailboxes, then used VPN configuration details recovered from mail to move deeper into internal networks.

The findings land as ARTEX appears to have spread rapidly among threat actors, prompting its developer, Li Puhua, also known as “Autumn-27”, to take the project closed source, according to Zscaler ThreatLabz. Zscaler said the framework can orchestrate several large language model back ends to carry out reconnaissance, find vulnerabilities and attempt exploitation with far less technical skill than would otherwise be required. CrowdStrike’s wider report also pointed to other China-linked groups it tracks, including Overcast Panda, which it associates with the FlowCloud backdoor and access through devices physically taken into China, and Titan Panda, which it says targeted legal billing software through a hard-coded IIS machine key flaw in Intapp Time Server to enable remote code execution.

Source Reference Map

Inspired by headline at: ^[1]

Sources by paragraph:

Source: Fuse Wire Services