Cybersecurity researchers have identified what is believed to be the first known example of malware that leverages a Large Language Model (LLM) to generate malicious code at runtime, known as MalTerminal.
See also: How Fileless Malware Differs from Traditional Attacks
Dubbed 'MalTerminal' by SentinelLABS, the malware uses OpenAI's GPT-4 to dynamically generate ransomware code and reverse-engineered packages, presenting a new and powerful challenge for threat detection and analysis.
The discovery highlights a significant shift in opponents' techniques, where malicious logic is not embedded in the malware itself but generated on the fly by an external AI model. This approach can render traditional security measures ineffective, such as static signatures, as the code can be unique for each execution. The findings were part of a broader research into how malicious actors use LLMs.
In contrast to other hostile uses of AI, such as creating persuasive phishing emails or using AI software as a lure, LLM-enabled malware embeds the model's capabilities directly into its payload. This allows the malware to adapt its behavior based on the target environment.
See also: Hackers use WhatsApp to distribute SORVEPOTEL malware
The researchers at SentinelLABS defined a clear definition for this threat, distinguishing it from malicious software that is simply generated by an LLM, which they note remains immature. The main concern with LLM-enabled malware is its unpredictability. By offloading code generation to an LLM, the actions of the malware can vary significantly, making it difficult for security tools to predict and block its behavior.
Previous documented cases such as PromptLock, a proof-of-concept ransomware, and LameHug (or PROMPTSTEAL), linked to the Russian group APT28, showed how LLMs could be used to generate system commands and extract data. These examples paved the way for the search for more advanced threats.
The discovery came from a new threat detection methodology developed by SentinelLABS. Instead of looking for malicious code, the researchers looked for the artifacts of LLM embedding: embedded API keys and specific prompt structures. They wrote YARA to detect key patterns for major LLM providers like OpenAI and Anthropic. A year-long retrospective search on VirusTotal flagged over 7,000 samples with embedded keys, though most were non-malicious developer errors.
See also: Detour Dog malware distributes Strela Stealer via DNS TXT Records
The key to finding MalTerminal was focusing on samples with multiple API keys, a redundancy tactic for malware, and looking for prompts with malicious intent. The researchers used an LLM classifier to score the malicious nature of the discovered prompts. This strategy led them to a set of Python scripts and a Windows executable named MalTerminal.exe.
☁️ Keep safe copies with Proton Drive
Encrypted cloud storage from Proton — protect your files from ransomware, corruption, and data loss with end-to-end encryption.
- ✔ End-to-end encrypted files & backups
- ✔ Version history — recover files after ransomware
- ✔ Free space — sync across all devices
The link is an affiliate link — SecNews may receive a commission at no additional cost to you. It does not affect the independence of our article writing.
