Sentinel is a proof-of-concept project developed during a one-day AI Agents hackathon. It showcases a novel approach to authorization systems designed specifically for AI Agents.

As AI Agents become more powerful and autonomous, traditional authorization systems need to be reimagined. Sentinel demonstrates a new approach to authz systems that caters to the unique needs of AI Agents, who are likely to be the primary consumers of software in the near future.

AI Agents are typically short-lived and operate across various tools and third-party interactions. Each agent's capabilities should be constrained to the minimum set of permissions required to perform its task. While complex APIs often allow third-party integrations through API Keys and OAuth tokens, many lack granular control over permissions. Even APIs with more advanced scoping (like GitHub's) may not be ideal for short-lived auth tokens required by AI Agents.

We've built a GitHub API wrapper with three endpoints:

• /create: Creates an issue in our repository
• /solve: Adds the solved tag to an issue
• /close: Closes an issue

The server implements a custom authorization system using Macaroons. Before processing a request, it checks two policies:

• Can the client execute the requested method?
• Are the credentials expired?

The client demonstrates the following workflow:

Generates an "admin" token with access to all three endpoints, expiring in 1 day. Creates two sets of credentials by narrowing the scope of the admin token:

• Set 1: Can only use the create and solve endpoints
• Set 2: Can use all endpoints but expires after 10 seconds

Spawns two agents, each with a different set of credentials.

• Agent 1 (with Set 1 credentials):

  • Creates two issues
  • Marks an issue as solved
  • Attempts to close an issue (which fails due to lack of permissions)

• Agent 2 (with Set 2 credentials):

  • Closes the first issue
  • Waits briefly
  • Attempts to close the second issue (which fails due to expired credentials)

This simple scenario demonstrates the possibilities of authorization systems built on the Attribute-based access control (ABAC) paradigm.

By providing fine-grained, time-bound permissions, Sentinel offers a glimpse into how we can better secure and control AI Agent interactions with APIs and services.