Can we lend 'permissions' to AI? The story of 'Pigeon', the signed Pass

Imagine this: You ask your personal assistant, “Organize today’s afternoon meeting materials and email them to the team members.” But suddenly, the assistant tries to access your bank account or posts content under your name on unauthorized external sites. It’s a chilling thought.

As we entrust more complex and sensitive tasks to AI agents (AI that decides and executes specific goals autonomously) in our daily lives, this ‘security problem’ has become a realistic concern. While it is important for AI to perform tasks intelligently, it has become even more crucial to safely control it so it accurately does only what we permit. Today, we introduce the ‘Pigeon’ protocol, a smart agreement that has emerged to solve this issue.

Why is security so important?

The AI we have mainly used so far involved inputting a single prompt (command) and receiving an answer. However, to assign complex tasks to AI—such as researching several competitors, analyzing that data, and writing a sophisticated report—the technology of ‘Sub-agents’ (lower-level AI delegated tasks from the main agent) is essential Source: Subagents: The Building Block of Agentic AI.

The problem is that when the Main AI delegates work to a Sub AI, it is very difficult to define the boundaries of how far that Sub AI is allowed to act. Pigeon clearly solves this ‘permission delegation’ problem. It works on the same principle as giving your assistant a very specific job description, saying, “Only copy these documents.”

A simple analogy

In short, the Pigeon protocol can be compared to a ‘digital work power of attorney.’

1. Scope of Permission (Pass): The main AI agent issues a ‘Pass’ to the sub-agent. This document details which resources the sub-agent can use, what actions it can take, and what it absolutely must not do Source: Pigeon, a signed Pass for what a sub-agent may do.
2. Pre-verification: Before the sub-agent actually starts working, the Pigeon system meticulously checks this ‘power of attorney.’ If it tries to do something you haven’t asked for, it is blocked from even starting Source: Pigeon, a signed Pass for what a sub-agent may do.
3. Strict Fail Closed Principle: What happens if the sub-agent asks for more permissions than it was granted or secretly tries to do something else? Pigeon decisively stops operation and fails the task Source: Pigeon, a signed Pass for what a sub-agent may do.

Simply put, when giving a ‘key’ to an AI, Pigeon is a meticulous safety device that hands over a ‘custom master key’ capable of opening only the necessary doors and immediately retrieves it if it attempts to open any others.

Current situation

In the current AI industry, task automation using sub-agents is progressing rapidly. Many development environments are already using sub-agents to perform coding tasks or analyze vast project data Source: Subagents - Docs by LangChain. However, due to a lack of unified security protocols, users are often anxious about how much permission to give to the AI.

A key feature of Pigeon is that it operates without going through a central server, allowing these security rules to be applied easily without complex server management Source: Pigeon, a signed Pass for what a sub-agent may do.

Future outlook

The AI assistants we use in the future will have even more autonomy. Beyond simply answering questions, they will manage our emails, adjust our schedules, and even handle sophisticated documentation. In this era, technologies like Pigeon will become the core standard for proving that ‘the AI is truly safe.’

As technology advances, AI’s judgment will become important, but keep an eye on these ‘invisible safety devices’ that help users safely entrust complex tasks to AI. Ultimately, what makes us trust AI and delegate tasks to it is these meticulous and strict agreements.

MindTickleBytes AI Reporter’s View

As the era of AI agents approaches, security must become a ‘basic’ requirement included from the design stage, not something to be ‘considered later.’ Technical attempts that force ‘least privilege,’ such as the Pigeon protocol, will accelerate a safer future where AI and humans coexist.

References

1. [Pigeon, a signed Pass for what a sub-agent may do

   Hacker News](https://news.ycombinator.com/item?id=49585209)

2. Subagents: The Building Block of Agentic AI - DEV Community
3. Subagents - Docs by LangChain