The use of artificial intelligence (AI) is reshaping various industries, including cybersecurity. Palo Alto Networks’ Unit 42 team recently conducted an analysis of 405 malware samples that were influenced by AI in some capacity. These ranged from ransomware partially developed with the aid of large language models (LLMs) to malware installers that merely adopted the name of popular AI applications.
AI’s Role in Malware Development
Unit 42’s investigation revealed that approximately 97% of these samples remained confined within sandboxes or research environments, never reaching actual targets. By cross-referencing file hashes against endpoint telemetry and network session data, the team discovered that only 12 instances had been detected on live systems, while a slightly larger number (15-20) appeared in network sandbox traffic. Each of these 12 samples triggered security alerts on protected endpoints.
The malware samples that did not make it to production were categorized into three groups. The largest group consisted of proof-of-concept code aimed at demonstrating specific techniques, often targeting local or private networks and filled with debug information. Another group included samples from organizations testing defenses against AI-driven malware, identified by repeated uploads from the same source. The third group used AI branding as a lure, disguising ordinary payloads as installers for known AI products without actual AI capabilities.
Detecting and Analyzing AI-linked Malware
The 12 samples that reached live environments were distributed among five malware families across three countries, with no particular industry or region being more affected. The most prevalent was FunkSec, a ransomware strain linked to LLM assistance, as evidenced by the iterative naming detected in internal project files—a pattern more aligned with prompt-driven generation than traditional development.
One notable sample, disguised as a recipe-finding app named Recipe Lister, managed to infiltrate over 50 organizations, resulting in around 6,500 endpoint records and about 9,600 alerts. Although its digital signature initially evaded detection, anomalies in the signer and heavily packed file contents eventually led to its identification. Another strain, the Oyster backdoor, impersonated a Dropbox installer, showcasing how attackers utilize AI tools to expedite and economize the creation of delivery code.
Implications for Cybersecurity Defense
Despite the AI involvement, existing defenses effectively intercepted all samples using conventional methods like sandbox detonation, behavior-based detection, and analysis of digital signatures. These findings suggest that AI currently serves more as a tool to accelerate the development of malware rather than enhance its evasion capabilities.
In conclusion, while AI is accelerating the rate at which attackers can innovate and diversify their tools, it does not necessarily make them more elusive. The cybersecurity community continues to rely on established detection techniques to counteract these threats, underscoring the importance of maintaining robust security measures.
