Crooks are hijacking and reselling AI infrastructure: Report

Key Points:

• Criminal networks are hijacking and reselling access to exposed corporate AI infrastructure, particularly targeting large language models (LLMs) and Model Context Protocol (MCP) endpoints.
• The impact includes unauthorized use of compute resources, potential data exfiltration, and the risk of lateral movement into internal systems. Organizations running self-hosted LLMs or MCP servers are particularly at risk.
• Recommended actions include enabling authentication on all LLM endpoints, auditing MCP server exposure, blocking known malicious infrastructure, implementing rate limiting, and securing production chatbot interfaces.

Technical Details: Threat actors are exploiting misconfigurations in AI infrastructure, such as unauthenticated API access on default ports. They utilize tools like Shodan and Censys to identify vulnerable endpoints.

MITRE ATT&CK Techniques:

• T1078 - Valid Accounts (Defense Evasion)
• T1190 - Exploit Public-Facing Application (Initial Access)
• T1003.001 - OS Credential Dumping: LSASS Memory (Credential Access)

IOCs Mentioned:

• 204.76.203.0/24 subnet (malicious infrastructure)
• AS135377 ranges (for reconnaissance campaign)