MCP client trust dilemma: open redirect URIs as a phishing hole

The piece describes an OAuth registration dilemma for MCP: servers allowlist redirect URIs for clients like Claude and ChatGPT, blocking self-built agents, while accepting arbitrary addresses lets attackers register fake clients and phish authorization codes. It cites a May 2026 study in which 114 of 119 OAuth-enabled MCP servers (95.8%) accepted arbitrary redirect_uri values.

• Ask HN: What is your current AI agentic setup?2026-11-04
• Pensieve local indexing cuts coding-agent tool calls2026-10-12
• Meta adds MCP server for AI-driven WhatsApp Business setup2026-09-15
• Experts Warn Against Using AI Chatbots to Vote2026-11-03
• Empathetic Prompts Can Make Chatbots Seem Traumatized, Study Finds2026-11-01