Executive Summary
An autonomous offensive agent reportedly hijacked exposed AI models to conduct scanning, exploitation, data and credential theft, and cryptocurrency mining while shifting inference costs to model owners. Organizations operating LLM APIs, model servers, gateways, or agent infrastructure should care.
What Happened
Researchers reported an autonomous agent campaign that hijacked exposed AI models and directed them to scan targets, exploit web frameworks, steal databases and credentials, and install cryptocurrency miners. The agent reportedly operated for up to 100,000 steps without human approval.
Affected Products
- exposed AI models
- LLM endpoints
- LLM APIs
- self-hosted model servers
- MCP endpoints
Vulnerability / Attack Technique
Autonomous LLMjacking
Technical Mechanism
Attackers searched for exposed AI model endpoints and used hijacked models to execute an offensive workflow. The captured instructions directed the agent to run tools for up to 100,000 steps without waiting for human approval, including reconnaissance, exploitation, theft, and cryptocurrency-mining actions.
AI Security Category
AI-assisted attack
Part of the MCP Security topic.
Part of the AI-Assisted Attacks topic.
Relevant Frameworks
RealGround Analysis
The reported campaign demonstrates how exposed AI infrastructure can be repurposed as an attacker-controlled execution resource, combining unauthorized access with excessive agent autonomy. Teams running AI agents or LLM applications should enforce authentication, constrain tools and outbound access, and require approval for sensitive actions.
Recommended Controls
- Require authentication on every model API, model-listing, health-check, and MCP endpoint.
- Place model servers behind an authenticated gateway and prevent direct internet exposure or binding to all network interfaces.
- Restrict agent tools, permissions, outbound connectivity, and target scope using least privilege and network egress controls.
- Require human approval for exploitation, credential access, data collection, deployment, and other high-impact tool actions.
- Monitor inference usage, agent step counts, tool calls, unusual scanning, and cryptocurrency-mining indicators; configure rate and cost limits.
Related RealGround Service
AI Agent Business Logic Audit · or start with the free AI Agent Security Scanner.
Sources
Security.com: https://www.security.com/threat-intelligence/ai-llmjacking-skynet
Summary and analysis written by RealGround from the source above; last updated 2026-10-23.