405 samples, twelve on endpoints

Sara McBroom published Unit 42's state of AI-enabled malware on 25 August. The team collected 405 unique samples that integrate AI capabilities, drawing on WildFire analysis reports, VirusTotal Intelligence and open-source research, and then asked the question the category usually skips: how many of these were ever on anyone's computer.

The answer

• 12 of 405 samples, three percent, were seen on endpoints.
• Around fifteen to twenty unique hashes, roughly four percent, appeared in production environments via firewalls.
• About 97% existed only in sandboxes and repositories.

The 97% breaks down into proof-of-concept and research code, security validation submissions, and a category worth naming on its own: malware branded as AI without any functional AI integration in it.

The five that made it out

• FunkSec ransomware, seven variants compiled between 1 and 6 January 2025.
• A trojanised Recipe Lister application, which reached more than 50 organisations.
• The Oyster backdoor, also tracked as CleanBoost.
• Rhadamanthys stealer, as a .NET executable.
• A COM hijacking DLL impersonating 360 Total Security.

The line that matters for a defender

Unit 42 states that existing behavioural detection, cloud sandboxing and endpoint analytics catch these threats through the same mechanisms that stop conventional malware. No new detection method was required.

That is a reason to keep spending on the controls already in place rather than on a product category named after the adversary's toolkit. It is not a reason to stop counting: three percent of 405 is a real number, and the trojanised application alone reached fifty organisations.