AI-Driven Espionage: Chinese APT Clusters Use LLMs to Automate Global Government Intrusions

Executive Summary Recent telemetry from Hunt.io and Encrygma Intelligence reveals a paradigm shift in cyber espionage tactics. Chinese state-sponsored threat actors have successfully integrated Large Language Models (LLMs) into their offensive workflows, effectively automating the lifecycle of intrusions against sovereign entities. This campaign, which has reached a peak in the last 48 hours, has compromised multiple high-value targets by utilizing AI to adapt exploits in real-time and craft highly personalized phishing lures. The discovery of an open directory containing thousands of operator logs provides a rare, unvarnished look at how nation-state actors are scaling their operations through machine intelligence. ## Threat Analysis The ongoing campaign primarily targets government administrative systems and financial institutions in Thailand, Taiwan, and the United States. Unlike traditional scripted attacks, this AI-assisted approach allows threat actors to rapidly iterate on failed exploits and reason through complex security bypasses. The strategic goal of the operation remains long-term intelligence collection on diplomatic negotiations, economic policies, and defense logistics. By leveraging LLMs to handle the 'heavy lifting' of exploit modification and code cleanup, Chinese clusters are now capable of maintaining a significantly larger volume of concurrent intrusions with fewer human operators. ## Technical Details The attackers utilized a sophisticated command-and-control (C2) infrastructure linked to the TencShell framework. Analysis of the logs found on a Hong Kong-based server revealed that operators were querying Anthropic’s Claude Code and DeepSeek-V3 models to automate the development of SQL injection strings and rework payload signatures after they were flagged by EDR systems. A notable development in this campaign is the resurfacing of the 'Daxin' kernel-mode rootkit alongside a new, stealthy backdoor dubbed 'Stupig.' Stupig trojanizes legitimate Windows keyboard-layout DLLs (such as kbdus.dll) to gain SYSTEM-level execution at the Windows logon screen before a user even signs in. This technique bypasses traditional logon audit events and provides a persistent, nearly invisible entry point for the attackers. ## Attribution Assessment This activity is attributed to a Chinese-nexus threat cluster, likely UAT-7810 or an advanced sub-group of APT41, with high confidence. The attribution is based on several key factors: the use of Simplified Chinese in the operator logs, the clustering of Hong Kong-based server infrastructure, and the specific deployment of the Daxin rootkit—a hallmark tool used by the Ministry of State Security (MSS) for over a decade. The precision and resource-intensive nature of using multiple high-end LLM APIs further suggest a well-funded, state-sponsored mission profile rather than an independent cybercriminal operation. ## Implications The successful integration of LLMs into state-sponsored workflows marks the transition to the 'AI-speed' attack era. Standard defense mechanisms and human-centric incident response teams are increasingly unable to match the speed at which AI can generate new variants of known threats. The focus on Taiwan’s manufacturing sector and U.S. government infrastructure indicates a coordinated effort to erode technological advantages and gain leverage in regional geopolitical disputes. Furthermore, the use of LLMs to automate reasoning for bypass techniques suggests that even 'fully patched' systems are at higher risk of being compromised by AI-discovered edge cases. ## Recommendations To counter these advanced threats, Encrygma recommends that organizations immediately move toward AI-enhanced behavioral monitoring and automated remediation platforms. Security teams should prioritize the auditing of Windows keyboard-layout DLLs and other winlogon-loaded components for unauthorized modifications. Additionally, it is critical to implement strict egress filtering on development environments to prevent the unauthorized use of commercial LLM APIs for malicious purposes. Zero-trust architecture remains the most effective long-term defense against lateral movement once an AI-driven initial breach has occurred.