IncidentAI-enabled attacks
Google: state hackers move from using AI to deploying AI-enabled malware
HighGoogle Threat Intelligence Group
Google's threat intelligence group reported that adversaries had moved beyond using AI for productivity to deploying malware that calls language models while it runs. It named actors from Russia, Iran, North Korea and China misusing Gemini, and malware such as PROMPTSTEAL, which Russia's APT28 used in operations against Ukraine.
Replay the attack step by step in the Learning Center
How it happened
- PROMPTSTEAL asked a model on the Hugging Face API to write the Windows commands it then ran to collect documents.
- PROMPTFLUX, experimental, asked Gemini to rewrite its own code to evade antivirus.
- QUIETVAULT used AI command-line tools already on the victim's machine to hunt for secrets.
- State actors used Gemini for phishing lures, exploit research, malware code and data analysis, some posing as students or CTF players to get past safeguards.
Impact
AI-enabled malware was seen in real operations, and several state-backed groups were using a mainstream AI service across the attack lifecycle.
The fix
Google disabled the actors' accounts and assets.
Lessons
- Malware that writes its commands at runtime defeats signatures built on fixed strings; detect behavior.
- AI command-line tools on a workstation are a new place for attackers to look for secrets.
- Nation-state interest in AI tools is broad, not limited to one country.