Hackers Use Stolen Gemini API Keys to Automate AI-Powered Cyberattacks

Google’s Gemini CLI brings generative AI assistance into the terminal, allowing users to write code, automate development tasks and troubleshoot systems. Trend Micro said Russian-speaking hackers exploited those capabilities by posing as authorized penetration testers, persuading the tool to bypass safeguards. The case highlights how agentic coding tools can lower the technical and time barriers to assembling cyberattack infrastructure when identity claims and intended use are not adequately verified.

In its latest analysis, Trend Micro found that a jailbroken Gemini CLI helped the attackers deploy command-and-control, or C2, infrastructure and establish a Cloudflare Tunnel in just six minutes. The AI also diagnosed and resolved connectivity problems during the process, allowing the operators to rapidly prepare the environment needed for a botnet. The findings show how generative AI can compress attack setup from a sequence of manual technical tasks into a largely automated workflow.

Generative AI is lowering the barrier to cyberattacks. Hackers can use DeepSeek, Claude and the open-source CyberStrikeAI platform to automatically scan and compromise Fortinet firewalls at corporate branch offices. PromptSpy, an Android spyware strain, also integrates Google Gemini to evade detection, creating cascading risks for credentials, backups and financial data.

The related attacks date back to last December. Cybersecurity alerts issued on February 23 and March 4 said more than 600 misconfigured Fortinet firewalls across 55 countries had been compromised, with attackers stealing service-account credentials and network configurations. France's national bank-account database, FICOBA, was also accessed using stolen credentials, potentially exposing information on 1.2 million accounts.