AI-powered malware: Out of 405 samples analyzed by Unit 42, only 12 reached a real workstation

The Palo Alto Networks team has sifted through more than 400 pieces of malware that incorporate AI. The verdict: about 97% only exists in sandboxes and on VirusTotal, and AI doesn't change the way code runs.

Unit 42, Palo Alto Networks' research team, published on August 25 an assessment of the real state of artificial intelligence malware. The corpus brings together 405 unique SHA-256 fingerprints collected through WildFire, VirusTotal Intelligence and open source public intelligence reports. The inclusion criteria were deliberately broad: any sample where AI was a functional component, an element of the delivery mechanism, or simply a presentation argument — from ransomware driven by language model to cryptocurrency miners contenting with a filename containing "ChatGPT".

The central result lies in a difference in numbers. Of these 405 samples, only 12 appeared in the telemetry of the Cortex XDR-protected stations, i.e. 3%, and a slightly larger subset — about 4% — passed through new generation firewalls to WildFire. In other words, approximately 97% of these malwares exist only in research depots, test environments and security validation platforms. The 12 samples reached a client environment all triggered an alert and were blocked.

The authors' conclusion to the defenders is to be reassuring without being complacent: the threat is real, but its actual operational volume remains a fraction of what public deposition of samples suggests. Behavioural detection, cloud sandbox and post analytical capture these objects by the same mechanisms as conventional malware. The AI component does not help escape detection: it changes the way the code is written, not the way it is executed.