Your AI can act. It shouldn't decide its own authority.
Autonomous software can reason, plan, browse, call tools, deploy code, move money, modify systems, and instruct other agents. None of that means it should be allowed to decide its own permissions.
Intelligence and authority are different things.
Models will make mistakes. They can also be manipulated by bad data, malicious webpages, compromised tools, or other agents.
The security failure is not that the model had a bad idea. The failure is allowing that idea to become a real-world side effect without an independent authority decision.
Put authority where intent becomes action.
Red Button sits at the execution boundary. The agent may request an action. The real tool only receives it when valid authority exists.
Prompt injection should not automatically become execution.
An agent can read hostile instructions. Red Button does not need to understand why the model changed its mind. It only needs to enforce whether the resulting action is authorized.
Authority can be explicit, bounded, and revocable.
Red Button is designed around a simple vocabulary. Authority may be standing, one-time, temporary, delegated, constrained, or revoked entirely. Human approval is one authority source—not the product category.
MCP is the first integration. The larger idea is portable runtime authority across the systems autonomous software can touch.
No side effect without authority.