How AI Is Changing the Malware Landscape
summary
VirusTotal deploys an LLM-based Code Insight over a subset of its ~50 billion-sample corpus and finds LLMs often explain scripts where AVs miss them: exploit samples flagged by LLMs are detected by 0–5 AVs for 69% of cases and by zero AVs for 41%. Agreement with AVs is 96.23% on PowerShell but only 72.4% on PHP; conventional tools catch ~5% of obfuscations that LLMs deobfuscate. Observations include prompt-injection attempts, a mixed-OS overclocking PowerShell prompt, and reliable file-type ID like LAMMPS inputs.
AI-written summary · about
chapters (13)
- 0:00AI and malware analysis
- 3:38VirusTotal Code Insight with LLMs
- 10:10LLMs detecting exploits
- 15:16LLMs vs antiviruses in malware
- 17:49LLMs handling obfuscated code
- 21:19LLMs identifying file types
- 24:52AI-driven malware tactics
- 28:10Detecting AI-generated malware
- 33:36AI in social engineering attacks
- 35:23AI lowering attacker bar
- 36:54AI changing malware?
- 37:59Malware detection implications
- 39:46Future of AI in malware research
related talks
36mCompromising LLMs: The Advent of AI Malware
36mUsing AI to Discover Silently Patched Vulnerabilities in Open...
27mA Machine Learning approach for detecting a Malware
09mHow Much Trust Do You Have with LLM-Based Solutions?
30mDeception at Scale: How Malware Abuses Trust
26mWhen Models Write Malware: Defending Against On-demand Exploits
39mLLMs everywhere: The future is now and the past keeps repeating itself
29mWorms, Tokens, and Trust The Industrialization of Supply Chain Attacks