The Model Context Protocol (MCP) enables secure AI connections to external data, crucial for AI-driven operations. As a new internet-facing infrastructure, it raises security concerns about potential attacks. GreyNoise conducted tests with MCP honeypots to monitor real interactions with AI middleware online. Three configurations were set up: unauthenticated endpoints for background scanning, authenticated endpoints to detect credential probing, and developer instances with exposed keys. All configurations were quickly discovered, confirming that online resources attract attention. Activity consisted of common internet probes, but no specific attacks targeting MCPs were found. A minor prompt-hijacking flaw was identified in a custom MCP, underscoring that risks stem from implementation errors rather than focused attacks. This quiet period signifies a baseline for AI middleware exposure. To enhance security, GreyNoise now offers an MCP server linked to its API for safe data analysis and dynamic blocking of suspicious activity. Visit the GreyNoise Labs blog for further insights.
Source link
