Anthropic MCP SDK Design Vulnerability

From The Hacker News:

https://thehackernews.com/2026/04/anthropic-mcp-design-vulnerability.html

There’s a design vulnerability in the Anthropic MCP SDK that allows carefully formed commands to cause the Agent to spawn processes that “fail successfully” (as in, the process is spawned, executes, but the SDK returns an error to the Agent.

Good News, Everyone

After looking through the attack models, and the Portcullis-Gate architecture, it’s pretty clear that this is not an issue for an agent using the open source PortcullisMCP authorization gateway to authorize and protect the MCP usage.

1. Portcullis-Gate does not use the process-spawning components of the SDK, so the command injection wouldn’t work if Gate is the single MCP available to the agent (which is what should be the case in any thoughtful enterprise security model)

2. Portcullis-Gate does not use the marketplace (also something that shouldn’t be done in a thoughtful enterprise security architecture)

3. Portcullis-Gate supports both stdio and http-based MCP calls, and this attack only applies to stdio calls, so a managed deployment of the Agent executable could lock this down from the start.