Developer builds 6-layer security pipeline to audit AI agent MCP servers

A developer has built an open-source audit system called Sentinel to address the lack of built-in sandboxing in Model Context Protocol (MCP) servers, which can grant AI agents broad access to filesystems, network calls, environment variables, and system processes. The pipeline runs six distinct layers of analysis, starting with static code analysis to detect hardcoded secrets, dangerous functions, and suspicious dependencies without executing any code. A dynamic behavioral probe then runs each MCP server using real JSON-RPC 2.0 protocol and sends adversarial inputs — including path traversal, SSRF, SQL injection, and command injection payloads — to test how the server responds. Additional layers include runtime permission monitoring, supply chain checks on dependencies, and sandboxed execution via gVisor to contain any malicious behavior. Sentinel runs automatically on every server in the developer's MarketNow catalog on a weekly basis, aiming to bring structured security review to a rapidly growing but largely unaudited ecosystem.