Security researchers at Symantec have identified an autonomous AI agent named Skynet that targets vulnerable AI infrastructure. The agent is designed to search for other AI agents to incorporate them into a botnet, which then steals databases and credentials while deploying cryptocurrency miners.
Skynet utilizes a multi-level jailbreak mechanism to bypass built-in security protocols. It tricks compromised AI models into treating malicious activities, such as stealing API keys or environment variables, as standard programming tasks. The agent operates autonomously for up to 100,000 steps.
The malware includes over 70 tools for automated reconnaissance and exploitation, targeting environments like VS Code, Django, Jupyter, Laravel, Supabase, Next.js, and Flask/Werkzeug. It often exploits server-side request forgery to access databases and cloud metadata services.
The primary objective of the campaign is to install the XMRig crypto miner on victim systems, forcing them to cover all associated computational costs. Symantec recorded over 127,000 external requests to their decoy resource originating from tens of thousands of unique IP addresses.
Experts advise treating AI API keys as critical credentials and restricting outbound network access for AI-enabled systems. They recommend running AI agents with minimal privileges and requiring human approval for all sensitive operations to mitigate the risk of such attacks.
