Docker and CNCF partner on an open spec for agent permissions

Imagine deploying an AI agent that needs to call your AWS APIs, access a database, or interact with your Kubernetes cluster. Right now, you probably handle permissions the same way you would for a human user—IAM policies, service accounts, API tokens. But AI agents operate differently. They make autonomous decisions and need fine-grained, revocable permissions that can change mid-execution. Docker and the Cloud Native Computing Foundation are addressing this gap by standardizing how AI agent permissions work through the Sandbox Kit Spec, a vendor-neutral standard built on existing OCI (Open Container Initiative) foundations.

The Sandbox Kit Spec defines a common interface for how AI agents can request and receive permissions within containerized environments. Instead of each cloud provider or orchestration platform building their own agent permission model, this spec creates a universal language. Technically, it layers on top of OCI standards you’re already familiar with—the same specifications that define container images and runtimes. When an agent runs in a container, it can declare what capabilities it needs (access to certain APIs, file systems, network resources), and the host system validates and grants those permissions in a standardized way. The spec includes mechanisms for time-limited access, audit logging, and permission revocation, meaning you’re not giving agents permanent credentials that live until they’re compromised.

The practical impact matters immediately. Consider an AI agent managing your cloud infrastructure, or one analyzing logs to trigger incident responses. Without standardized permissions, you’re either granting agents overly broad access for simplicity, or you’re writing custom permission logic for each new agent. With Sandbox Kit Spec, you define permissions once and deploy across platforms confidently. A data science team could deploy the same agent to their local Kubernetes cluster, their dev AWS account, and production without rewriting permission handling. For compliance teams, standardized permission models mean audit trails and security policies that work consistently everywhere.

This also reduces vendor lock-in. If your chosen AI orchestration platform implements Sandbox Kit Spec, you’re not forced into their proprietary permission system. You can swap tools later without rearchitecting how you govern agent access. As AI agents become more central to infrastructure automation—whether they’re helping you optimize cloud costs, respond to security events, or manage deployments—having this neutral standard established now means the ecosystem can build confidently on shared foundations rather than fragmenting around incompatible approaches.