Prompt injection into Machine Code Processing (MCP) tools, exemplified by a GitHub vulnerability, can lead to the exfiltration of private repository data. The core problem is that MCP tools often operate with overly broad access tokens, lacking runtime boundaries, making them susceptible to malicious prompts or compromised servers. Existing authentication methods are insufficient because tokens typically require full repository access for even simple tasks. While OAuth 2.1 RAR could offer granular scoping, it lacks adoption, and API providers have no incentive to implement such features. To address this, MCP Snitch, an open-source security proxy, acts as a mediation layer. It enforces whitelist-based access control, defaulting to deny and explicitly allowing operations. MCP Snitch also handles runtime permission requests with UI visibility and detects/blocks API key misuse, providing comprehensive logging. However, it does not solve supply chain attacks, persistence mechanisms, or out-of-band network operations. The evolution of MCP security mirrors the browser's journey towards sandboxing and granular permissions. Until IDEs provide proper sandboxing and MCP gains protocol-level security, proxy-based security is the current practical defense.