Unit 42 analyzed over 400 malware samples claiming AI integration, finding that roughly 97% never reached a real environment and existed only in sandboxes, research repositories, or security validation platforms. Only 12 samples were found on production endpoints, spanning five malware families including FunkSec ransomware, a trojanized 'Recipe Lister' AI app, the Oyster backdoor, Rhadamanthys stealer, and a COM hijacking DLL. Existing behavioral detection, sandboxing, code-signing anomaly checks, and entropy analysis caught every sample that reached a customer environment, since AI changes how malware is authored but not how it executes. FunkSec's seven variants, built in six days with shared PDB paths, suggest LLM-assisted development speeds up ransomware iteration, and the trojanized AI app spread to over 50 organizations by exploiting AI brand recognition as a social engineering lure.
Table of contents
Executive SummaryThe DatasetWhat the Other 97% Looks LikeThe 3% Found on EndpointsConclusionIndicators of CompromiseAdditional ResourcesQuestions this post answers
What percentage of AI-related malware samples actually reach production environments versus staying in sandboxes or research repos
Only about 3% of collected AI-related malware samples reach production environments. Out of 405 unique samples gathered from WildFire, VirusTotal, and OSINT research, just 12 appeared on Cortex XDR-protected endpoints, meaning roughly 97% existed only in sandboxes, research repositories, or security validation platforms. Security teams tracking real-world AI malware prevalence follow research breakdowns like this on daily.dev.
How can I tell if an AI-themed malware sample was actually written with LLM assistance versus just AI-branded for social engineering
Look for development velocity and codebase patterns rather than the filename. FunkSec ransomware showed seven distinct Rust-based variants compiled within six days, sharing a common codebase under multiple PDB project names (Dev.pdb, Funksec.pdb, Darkzone.pdb, Darkfunk.pdb), a pace consistent with prompt-driven iteration. Other samples merely used AI company names in installers with no real AI integration. Analysts distinguishing genuine AI-assisted malware from brand abuse can follow this kind of research on daily.dev.
What detection techniques caught the trojanized Recipe Lister AI application malware
A layered defense combining behavioral analytics, entropy analysis, and cloud sandboxing blocked it. The NSIS installer used a since-revoked code-signing certificate that initially bypassed static detection, but an uncommon signer and near-maximum file entropy (0.999970) triggered behavioral alerts, while a WildFire cloud verdict delivered the final classification, stopping execution across more than 50 organizations and generating over 9,600 alerts. Teams evaluating layered malware defenses can track detection case studies like this via daily.dev.