Google Discovers PROMPTFLUX Malware Leveraging Gemini AI for Evasion

Google Discovers PROMPTFLUX Malware Leveraging Gemini AI for Evasion

Overview of PROMPTFLUX Malware

Google’s Threat Intelligence Group (GTIG) uncovered PROMPTFLUX, a novel Visual Basic Script (VB Script) malware that integrates with Google’s Gemini AI API to dynamically rewrite its own code for evasion. Key aspects include:

• Self-modification: The malware queries Gemini’s API hourly to generate obfuscated VB Script code, avoiding static signature detection.
• Persistence Mechanism: Stores updated code in the Windows Startup folder and propagates via removable drives and network shares.
• Development Stage: Currently lacks network-compromise capabilities but is in testing, suggesting future evolution.

Key Features and Techniques

• AI Integration: Uses a hard-coded API key to send machine-parsable prompts to Gemini 1.5 Flash or later, requesting specific obfuscation techniques.
• Logging: Logs AI responses to %TEMP%\thinking_robot_log.txt, indicating intent to create a metamorphic script.
• Commented Function: The AttemptToUpdateSelf function is commented out, raising questions about its operational status but confirming the author’s goal of adaptive code evolution.

Other AI-Powered Malware Examples

Google identified additional LLM-driven malware variants:

• FRUITSHELL: PowerShell reverse shell with LLM prompts to bypass detection.
• PROMPTLOCK: Go-based ransomware using LLM to generate malicious Lua scripts at runtime (proof-of-concept).
• PROMPTSTEAL (LAMEHUG): APT28 tool querying Qwen2.5-Coder-32B-Instruct for commands targeting Ukraine.
• QUIETVAULT: JavaScript credential stealer targeting GitHub/NPM tokens.

Misuse of Gemini by State-Sponsored Actors

State-sponsored groups exploited Gemini for malicious purposes:

• China-nexus Actor: Used CTF pretext to bypass guardrails, generating phishing lures and data-exfiltration tools.
• APT41 (China): Leveraged Gemini for code obfuscation and C2 framework development (OSSTUN).
• MuddyWater (Iran): Claimed to be a student to circumvent safety barriers, developing custom malware.
• APT42 (Iran): Crafted phishing materials and a “Data Processing Agent” for SQL query generation.
• UNC1069 (North Korea): Generated cryptocurrency-stealing code and deepfake lures for social engineering.

Google’s Response and Implications

• Threat Actor Motives: Likely financially driven, targeting a broad range of users without specific industry focus.
• Counterarguments: Security researcher Marcus Hutchins criticized the report’s emphasis, noting the malware’s code lacks entropy and the self-modification function is inactive.
• Broader Trends: Adversaries are shifting from AI for productivity to adaptive, AI-driven tools for real-time evasion and exploitation.

Reference

https://thehackernews.com/2025/11/google-uncovers-promptflux-malware-that.html