ReversingLabs Finds Growing AI Signals in Malware, but No Measurable Attack Surge

Summary

ReversingLabs found more malicious files referencing LLM APIs, but says those references do not prove model use or autonomous attacks. Its research also examines AI-related malware, PHP2Shell samples and defensive priorities.

Key points

Malicious-file searches found 596 API-reference matches in 2024, 3,044 in 2025 and 10,821 through October 24, 2026; an API reference does not confirm model execution or attacker intent.
ReversingLabs estimated its higher-confidence named malware subset grew from about 150 files in 2024 to 620 in 2025 and 1,680 through October 24, 2026; these are estimates, not verified counts of AI-using malware.
The article describes PromptFlux, LameHug/LameHubloader and PromptSpy as examples of reported AI-related malware behavior, while noting ReversingLabs did not independently attribute its samples to actors.
ReversingLabs identified 10 PHP2Shell exploit files classified under CVE-2026-109618, first seen on disclosure day or within three days; 5%–12% of antivirus engines detected them, and the files do not demonstrate successful compromises.
The article recommends inspecting software artifacts, rapidly mitigating critical internet-facing vulnerabilities, monitoring behavior and unexpected model-API requests, and limiting AI tools’ credentials and privileges.

Article Details

Publisher: ReversingLabs
Report Period: 2024-01-01 to 2026-10-24
Scope: ReversingLabs malicious-file corpus searches for LLM API references, observations of reported AI-enabled malware, and PHP2Shell exploit files.
Sample Size: A pooled random sample of 160 named detections informed the higher-confidence estimates; corpus counts are not a prevalence denominator.
Key Statistics: Estimated higher-confidence files containing LLM API references: approximately 150 in 2024, 620 in 2025, and 1,680 from 2026-01-01 through 2026-10-24. These are estimates, not independently verified counts of AI-using malware.
The broader malicious-file search returned 596 matches in 2024, 3,044 in 2025, and 10,821 through 2026-10-24.
The 10,821 files found in 2026 contained 19,601 endpoint references, about 1.8 per file; endpoint presence does not prove model execution.
ReversingLabs identified 10 PHP2Shell exploit files classified under CVE-2026-109618, first observed on disclosure day or within the next three days; 5%–12% of antivirus engines detected them. Some may have been incomplete PoCs, not successful compromises.
VulnCheck reported 14 confirmed exploited vulnerabilities among 1,061 attributed to AI-assisted discovery in the first half of 2026, approximately 1.3%; the article cautions this is not proof of a higher exploitation rate.
Recommendations: Inspect delivered software artifacts, downloaded packages, and files before deployment or execution, including embedded components in third-party software.
For critical, internet-facing RCE vulnerabilities with credible exploitation, begin emergency mitigation immediately and prioritize patching or compensating controls within 24 hours where feasible; validate exposure in parallel.
Use behavioral signals alongside hashes and strings; monitor suspicious process activity and unexpected outbound model-API requests, including API routers.
Limit AI agents and integrations to necessary credentials, network access, and execution privileges; review unfamiliar tools and packages, monitor key usage, and prepare to revoke credentials quickly.