MCP security is still basically unsolved

A Show HN post for a real-time MCP interceptor that blocks .env reads and dangerous agent commands got a pointed response: 'If you want real security you do default deny not default allow.' The project's claimed public GitHub also returned a 404, which did not help the reception.

The pattern here is that MCP (Model Context Protocol, the standard for giving AI agents access to tools and filesystems) is getting adopted fast and secured slowly. The interceptor approach, blocking specific known-bad commands, is a denylist strategy. Security people on HN are allergic to denylists for exactly the reason the commenter named: you cannot enumerate all the bad things an agent might do.

This is a recurring theme in AI agent security discussions. Agents that can read files, run commands, and call APIs need a fundamentally different security model than web apps, and the tooling is not there yet.