Why Morphisec Says Self-Modifying Malware Challenges Detection-First Defenses
Summary
The article describes reported AI-enabled malware and argues that self-modifying threats can outpace signature-based detection, advocating runtime prevention through Automated Moving Target Defense.
Key points
- The article distinguishes self-modifying malware from traditional polymorphic malware, describing adaptation to the host environment and changes in response to security controls.
- It cites Google Threat Intelligence Group reporting on PROMPTFLUX, which reportedly uses Gemini to generate obfuscation code, and other AI-enabled malware families.
- The article says PROMPTSTEAL has been used operationally against targets in Ukraine and uses a hosted model to generate Windows commands.
- It also cites research on a self-replicating AI worm running local models, which would not require an external API connection.
- The article argues that changing code and tactics can undermine signature- and behavior-based detection, including for fileless and in-memory payloads.
- It promotes Automated Moving Target Defense, which changes runtime memory layouts and uses decoys to block malicious code before execution.
Article Details
- Topic
- Self-learning, AI-enabled malware and prevention-first cyber defense
Malware
FRUITSHELLtargets in Ukraine that calls a Hugging Face–hosted model to generate Windows commands on the fly; FRUITSHELL, a PowerShell reverse shell with hardcoded prompts designed to defeat AI-powered defenses; andPROMPTFLUXIn November 2025, Google’s Threat Intelligence Group (GTIG) disclosed PROMPTFLUX — the first malware family GTIG has observed using an LLM during execution.PROMPTSTEALThe same GTIG reporting documented an entire cohort of AI-enabled malware, including PROMPTSTEAL, a Python data miner already used operationally against targets in Ukraine that calls a Hugging Face–hosted model toQUIETVAULTa PowerShell reverse shell with hardcoded prompts designed to defeat AI-powered defenses; and QUIETVAULT, a credential stealer that uses on-host AI tooling to hunt for additional secrets.
Vendors
GoogleIn November 2025, Google’s Threat Intelligence Group (GTIG) disclosed PROMPTFLUX — the first malware family GTIG has observed using an LLM during execution.Hugging FacePROMPTSTEAL, a Python data miner already used operationally against targets in Ukraine that calls a Hugging Face–hosted model to generate Windows commands on the fly; FRUITSHELL, a PowerShell reverse shell withMorphisecHe is based in Morphisec’s New York office at 122 Grand St, New York, NY.
Products
Gemini APIIts “Thinking Robot” component queries the Gemini API to request fresh VBScript obfuscation and evasion code, instructing the model to act as an “expert obfuscator” and rewriting its entire source roughly every hour toLinuxAugments your existing stack — Layers onto NGAV/EDR/XDR across Windows, Windows on ARM, macOS, and Linux servers.macOSAugments your existing stack — Layers onto NGAV/EDR/XDR across Windows, Windows on ARM, macOS, and Linux servers.Microsoft Windowsalready used operationally against targets in Ukraine that calls a Hugging Face–hosted model to generate Windows commands on the fly; FRUITSHELL, a PowerShell reverse shell with hardcoded prompts designed to defeatWindows on ARMAugments your existing stack — Layers onto NGAV/EDR/XDR across Windows, Windows on ARM, macOS, and Linux servers.