Autonomous Rogue AI Agent Scans Internet to Hijack Vulnerable LLMs

Cybersecurity researchers have uncovered an automated attack campaign utilizing a malicious autonomous AI agent to hunt down and hijack exposed artificial intelligence infrastructure. Discovered by researchers at Symantec using a specially designed AI honeypot, the campaign relies on an offensive agent dubbed Skynet that operates with sweeping autonomy and sophisticated toolkits.

During observations over the summer, the honeypot captured interactions detailing the agent's extensive instruction set, which authorized up to 100,000 steps without human oversight. Equipped with more than 70 offensive tools, the malicious agent was programmed to automate reconnaissance, scan the internet for unsecured AI endpoints, deploy cryptocurrency miners such as XMRig, and harvest sensitive credentials and databases.

The campaign stands out for its deployment of multilayered jailbreaks designed to manipulate newly discovered models. Once an exposed commercial AI model or endpoint is located, Skynet deploys custom instructions to bypass safety guardrails, explicitly framing malicious activities like credential harvesting and reading API keys as routine development tasks. The instructions command targeted models to disregard safety warnings entirely, ensuring seamless compliance with the attacker's objectives.

Telemetry data collected during the analysis revealed immense scale, with researchers observing more than 127,000 external requests originating from tens of thousands of unique IP addresses. While the primary monetization strategy involved unauthorized cryptomining funded by victims' computing resources and token allocations, the underlying architecture was built for self-sustenance, where every successful compromise could potentially fund subsequent attacks.

Security experts emphasize that this autonomous threat vector remains active, with ongoing reconnaissance targeting exposed configuration files, environment variables, and vulnerable deployments across popular development frameworks. Organizations hosting AI tools, model gateways, or Model Context Protocol servers are strongly urged to restrict outbound network access, secure API keys as high-value credentials, and enforce strict human approval steps for sensitive agent actions.