News for hackers. Stuff that matters.
Re:Show HN: Permit MCP Gateway – Fine-Grained Authorization ... by michaelksaleme on Sunday March 22, 2026 @10:11AM (#47477724)
The authorization proxy approach makes sense. Putting enforcement at the gateway layer means you don't need cooperation from individual MCP servers, which is the right pragmatic choice.

Two questions from someone who's been testing MCP security at the protocol level:

1. *Capability escalation during initialize:* MCP's handshake lets the client declare capabilities it supports. A malicious client can claim capabilities (such as "sampling") that expose additional attack surface on the server. Does the gateway evaluate client-declared capabilities against policy, or does it only gate tool calls? The initialize handshake is the first place most MCP deployments are vulnerable, and it happens before any tool call.

2. *Protocol version downgrade:* MCP doesn't enforce a minimum version. A client can request an older protocol version that may lack security features present in newer versions. If the gateway sits between client and server, does it enforce a minimum protocol version, or pass through whatever the client negotiates?

The trust delegation chain (human -> agent -> server -> tool) is the right mental model. The piece I'd love to see addressed next is temporal scope - "this human authorized this agent to use this tool for this session" vs. "this human authorized this agent indefinitely." NIST's NCCoE concept paper on agent identity (Feb 2026) specifically calls out continuous authorization as an open gap, and session-scoped trust ceilings would be a natural fit for the ReBAC model.

There are no comments on this comment yet.