A new, sophisticated form of malware dubbed LAMEHUG is leveraging artificial intelligence to launch attacks on Windows systems, signaling a potentially dangerous new era in cyber warfare. Discovered by the Computer Emergency Response Team of Ukraine (CERT-UA), this malware is the first of its kind to be publicly documented to integrate a large language model (LLM) to dynamically generate malicious commands.
Ukrainian authorities have attributed the attack, with moderate confidence, to APT28, a group linked to the Russian military intelligence agency (GRU) and also known by various names such as Fancy Bear, Sednit, and Sofacy Group. APT28 has been active since at least 2004 and has a long history of targeting Ukraine with cyber-attacks.
LAMEHUG is primarily distributed through phishing campaigns. On July 10, 2025, CERT-UA identified a campaign targeting executive authorities with a ZIP file posing as a ministry document. The archive contained the LAMEHUG malware disguised as a .pif file, which is an executable file built in Python and packaged using PyInstaller.
What makes LAMEHUG particularly innovative and dangerous is its use of an LLM. Specifically, the malware uses Qwen 2.5-Coder-32B-Instruct, an open-source large language model developed by Alibaba Cloud, via the huggingface.co service's API. This LLM is optimized for coding tasks and can convert natural language descriptions into executable code or shell commands. This capability allows threat actors to adapt their tactics during a compromise without needing new payloads, potentially making the malware harder to detect by security software.
Once it infects a system, LAMEHUG executes system reconnaissance and data theft commands. It gathers basic information about the computer, including hardware, processes, services, and network connections, and stores it in a text file. It then recursively searches for Microsoft Office documents, PDFs, and TXT files in common folders like 'Documents,' 'Downloads,' and 'Desktop'. The collected data is then exfiltrated via SFTP or HTTP POST requests to attacker-controlled servers.
The emergence of LAMEHUG represents a significant evolution in cyber threats. The use of dynamically generated commands can help the malware remain undetected by security software or static analysis tools that look for hardcoded commands. Furthermore, leveraging Hugging Face's infrastructure for command and control purposes may help with making communication stealthier, keeping the intrusion undetected for a longer period. Security experts warn that this could usher in a new attack paradigm where threat actors can adapt their tactics during a compromise without needing new payloads.
While CERT-UA did not state whether the AI-generated commands were successfully executed, the very existence of LAMEHUG serves as a stark warning. As AI becomes more accessible, the potential for threat actors to weaponize it for malicious purposes grows. This development underscores the critical need for continuous innovation in defensive cybersecurity strategies to counter these evolving threats.