Google sounds alarm on self-modifying AI malware

Google’s Threat Intelligence Group (GTIG) warns of a new generation of malware that uses AI during execution to mutate, adapt, and collect data in real time, boosting evasion and persistence. In 2025, GTIG identified the first malware that mid-execution changes its behaviour, with early examples including PROMPTFLUX and PROMPTSTEAL that leverage large language models (LLMs) during execution to generate malicious scripts and commands on demand.

PROMPTFLUX is a VBScript dropper that uses the Gemini API to rewrite its own code and enable self-regeneration, while PROMPTSTEAL is a data miner that queries an LLM via Hugging Face to generate system-information and document-collection commands. Also noted were other AI-enabled tools such as FRUITSHELL, a PowerShell reverse shell, and QUIETVAULT, a JavaScript credential stealer, that exfiltrate data and tokens while tapping on AI prompts for evasion.

The report, according to Google, also highlights how state-sponsored actors from North Korea, Iran, and the PRC misuse generative AI tools to enhance reconnaissance, phishing, and data exfiltration, and confirms that Google has disabled associated assets and tightened model protections.