openai/openai-guardrails-python

Add minimal Pre-Action Authorization example pattern

Opened this issue · 0 comments

uchibeke commented

If this repo is the canonical place for guardrail patterns, I can contribute a minimal pre-action authorization example that demonstrates the pattern. The example is framework-agnostic (no OpenAI or runtime-specific code) and uses APort as the implementation example. Developers can adapt it to use their own authorization service.

APort Pre-Action Authorization Architecture

flowchart TB
    subgraph "Agent Application"
        A[User Request] --> B[Agent Runtime]
        B --> C{Action Decision}
        C -->|Wants to execute tool| D[Action Guardrail Wrapper]
    end

    subgraph "APort Service (Pre-Action Authorization)"
        D -->|Verify Request| E[APort API]
        E --> F[Load Passport<br/>Agent Identity & Limits]
        E --> G[Load Policy<br/>Rules & Requirements]
        F --> H[Policy Evaluator]
        G --> H
        H --> I{Evaluation Result}
        I -->|Pass| J[ALLOW<br/>decision_id emitted]
        I -->|Fail| K[DENY<br/>reasons provided]
    end

    subgraph "Tool Execution"
        J --> L[Tool Executes<br/>Side Effects Happen]
        K --> M[Tool Blocked<br/>No Side Effects]
        L --> N[Response]
        M --> N
    end

    style H fill:#2e7d32,stroke:#1b5e20,stroke-width:2px,color:#fff
    style J fill:#388e3c,stroke:#1b5e20,stroke-width:2px,color:#fff
    style K fill:#c62828,stroke:#b71c1c,stroke-width:2px,color:#fff
    style L fill:#1565c0,stroke:#0d47a1,stroke-width:2px,color:#fff
    style M fill:#e65100,stroke:#bf360c,stroke-width:2px,color:#fff
Loading

Key Point: Pre-action authorization runs after the LLM/agent decides what action to take but before any side effects occur. This addresses a distinct concern from input/output guardrails:

  • Input/output guardrails: Protect against malicious/unsafe data
  • Pre-action authorization: Enforce business policies, identity, and limits on actions

The proposed example demonstrates this as a generic pattern, using APort as the implementation example.

I have more details in the Agent Repo at openai/openai-agents-python#2022

Links and references