AI agents are increasingly capable of performing real-world operations:

- executing shell commands - calling APIs - manipulating files - interacting with infrastructure

However most existing authorization models are identity-based:

User → Permission → Action

AI systems do not behave like identities.

They behave like evolving execution processes driven by model reasoning.

During execution an AI system may:

- dynamically generate new tool calls - compose multiple capabilities - trigger external side effects - expand behavior through reasoning loops

This creates a mismatch between traditional authorization models and AI execution.

This repository proposes an alternative model:

Execution Context Authorization Model.

The key idea is that authorization should be bound to an *Execution Context*, not an identity.

The model defines:

- capability ceilings - capability requests - validation relations - external event authorization

Repository:

https://github.com/Madongming/context-capability-model

The goal is to define a *formal authorization model for AI execution* rather than a framework or implementation.

Curious to hear feedback from people working on:

- AI agent systems - capability-based security - runtime sandboxing - AI safety

Happy to answer questions about the model if anything is unclear.