Reference Architecture for AI Agent Security Released

The Agent-Safe Pipeline reference architecture has been released to address security risks (such as prompt injection and faulty judgment) that arise when granting direct execution privileges to AI agents.

The core design principle is "Let agents propose. Let policy decide." Agents are designed to only propose actions, without directly holding execution privileges or credentials.

Key Components and Flow:

• IntentCapture: Freezes and normalizes agent proposals into Immutable Intent.
• DecionisGate: Determines whether to ALLOW, ESCALATE, or BLOCK via an independent decision service.
• PresenceApprovalCoordinator: Requires human verification, but the approval result is used only as 'evidence' passed to the decision service, not as execution privileges.
• SafeExecutor: Executes registered handlers after verifying the validated intent and permissions.

Core Security Strategies:

• Privilege Boundary Separation: Credentials are placed behind the SafeExecutor rather than in the agent runtime, fundamentally preventing agents from directly calling subsystems or hijacking privileges.
• Threat Model: Designed on the premise that agent prompts, tool arguments, and model outputs are all untrusted.

This project is implemented in TypeScript and released under the Apache 2.0 license, allowing it to be used to build secure execution structures for AI agents in various environments.