mirage-ogpp

To enforce a single trusted gateway for untrusted agent runtimes, route high-level intents through the mirage_act tool. This centralizes control by preventing direct side-effect access and applying governance, sealed-handle enforcement, and dual-authorization execution.

Safety guardrails for agent side effects work by restricting runtimes to a single vetted tool. This mechanism applies sealed-handle enforcement and 2-server PIR-based policy checks to ensure dual-authorization before executing actions like ReadFile or SendMessage.

You route high-level intents through a gateway with privacy and governance controls by invoking the mirage_act tool. The gateway processes the intent, applies 2-server PIR checks for privacy, and returns an observable JSON response detailing the execution result.

Yes, the gateway approach requires dual-authorization for agent runtime execution. It enforces this alongside sealed-handle models and 2-server PIR-based policy checks to ensure untrusted runtimes cannot perform side effects without strict verification.

The best way to prevent direct side-effect access from untrusted agents is exposing a single vetted tool as the sole interface. Centralizing intent routing through this gateway applies necessary safety guardrails and prevents opaque data exposure.