Human-in-the-loop (HITL)

A checkpoint on actions that are irreversible, costly, or outward-facing: sending, publishing, deleting, paying, granting access. The agent proposes; the human confirms with enough context to judge. MCP's design assumes this for tool calls.

Why it matters

It is a strong defense that degrades badly. Done well — rare prompts, clear stakes, dangerous requests visually distinct — it catches the exfiltration attempt a filter missed. Done poorly — constant identical dialogs — it trains users to approve everything and provides only the appearance of oversight (see consent phishing). The quality of the UX is the control.