Unit 42 Finds AI Accelerates Malware Development More Than Its Impact

Artificial intelligence appears to be helping threat actors create and modify malware more quickly, but a recent analysis suggests it has not yet produced a significant increase in successful infections.

Palo Alto Networks’ Unit 42 examined 405 malware samples associated with AI. The dataset included malicious software reportedly developed with assistance from large language models, as well as conventional payloads disguised as installers for popular AI applications. Researchers compared the samples with endpoint telemetry, sandbox network activity and internal security alerts.

Only 12 of the 405 file hashes were observed on live, protected endpoints. A larger number—approximately 15 to 20—appeared in sandbox traffic. Each of the 12 samples found on production systems generated a security alert, leading Unit 42 to conclude that about 97% of the samples did not progress beyond research, testing or isolated analysis environments.

What the sample set contained

Unit 42 grouped most of the non-production samples into three categories. Some were proof-of-concept projects designed to demonstrate techniques and included testing artifacts unlikely to appear in operational malware. Others were repeatedly uploaded by organizations testing defensive controls. A third category used AI-related branding as social-engineering bait, despite containing no meaningful AI functionality.

The 12 samples seen on endpoints represented five malware families across three countries. FunkSec ransomware was the most frequently identified family and showed development patterns that researchers said were consistent with rapid, prompt-assisted iteration.

One particularly widespread sample masqueraded as a recipe application called Recipe Lister. Although digitally signed, it launched a backdoor after installation and appeared across more than 50 organizations, producing roughly 6,500 endpoint records and 9,600 alerts. Other examples included an Oyster backdoor disguised as Dropbox, a Rhadamanthys information-stealing payload, and malware impersonating a component of 360 Total Security.

Conventional defenses remained effective

Unit 42 reported that existing controls detected the samples through sandbox execution, behavioral analysis, suspicious signing information and indicators such as heavy packing or encryption. None required a new detection technique.

The findings indicate that AI’s current effect on malware is primarily operational: it can reduce development time and make variation cheaper, while established security technologies remain capable of identifying the resulting threats.