MCP-Security-Framework

The framework employs strong sandboxing via Docker containers to isolate target MCP servers and their dependencies. All interactions with targets are mediated by a SafeAdapter, enforcing scope, rate limits, and evidence redaction. While executing untrusted code (dependency installation, server startup) is inherent to its testing function, this occurs within an isolated Docker environment. The crash analysis loop, which may install system dependencies (e.g., via `apt-get install`), broadens the in-container attack surface but remains confined within the sandbox. No direct hardcoded secrets for framework operation or obfuscated code were detected in the provided source code. The design prioritizes non-destructive assessment of external targets.