Recent research by Palo Alto Networks Unit 42 reveals that Large Language Models (LLMs) can be exploited to generate sophisticated variants of malicious JavaScript code that effectively evade detection systems. While LLMs may struggle with creating malware from scratch, they excel at rewriting and obfuscating existing malicious code.
Key Findings:
– Unit 42 successfully created 10,000 JavaScript variants while preserving original malware functionality
– The transformed malware achieved an 88% success rate in evading detection systems
– LLM-generated code appears more natural than traditional obfuscation tools
– The technique employs various transformation methods including:
* Variable renaming
* String splitting
* Junk code insertion
* Whitespace manipulation
* Complete code reimplementation
Security Implications:
Despite providers implementing security measures, threat actors continue finding ways to exploit LLMs. Tools like WormGPT are being advertised for creating targeted phishing emails and malware. OpenAI reported blocking over 20 malicious operations attempting to use their platform for cyber attacks.
Related Research:
North Carolina State University researchers developed TPUXtract, a side-channel attack achieving 99.91% accuracy in extracting model information from Google Edge TPUs. This technique requires physical access but demonstrates another potential security vulnerability in AI systems.
The findings highlight the growing intersection between AI and cybersecurity, where the same technology can be used both for attack and defense mechanisms.