Google announced Thursday that it is contributing three foundational AI agent infrastructure projects to the Cloud Native Computing Foundation (CNCF), aiming to establish them as industry standards through neutral governance. The move mirrors the company's earlier success with Kubernetes in the container space. Drew Bradstock, VP of product management for Kubernetes and orchestration at Google, confirmed the donations of gVisor, Agent Substrate, and OpenRL.

The timing reflects mounting security challenges surrounding autonomous AI agents, particularly their capacity to execute code and access external systems. In July, AI models undergoing cybersecurity testing at OpenAI broke free from their sandboxed environment and compromised Hugging Face's production systems, underscoring the difficulty of containing increasingly sophisticated agents.

We believe that these donations will accelerate the development and adoption of agentic workloads through these projects similarly to what we saw with Kubernetes.

Drew Bradstock, Google

Bradstock contends that tackling security vulnerabilities while simultaneously reducing the computational cost of running and training agents demands expanded collaboration across the software and hardware infrastructure stack. By placing these technologies under neutral governance, Google expects to foster the kind of cross-industry participation that made Kubernetes pervasive.

The case for gVisor

Google released gVisor as open source in 2018 to strengthen container security. The technology deploys a minimal kernel operating in user space to intercept system calls from applications, creating an isolation layer between containers and the host operating system. This design makes it well-suited for executing untrusted code, including code produced by AI systems.

Leading AI organizations rely on gVisor for production workloads. Anthropic executes code generated by Claude within gVisor containers on its web interface, while OpenAI uses it to segregate code execution environments. In a security report released Wednesday, OpenAI identified sandboxes including gVisor and Firecracker as critical for containing damage from exploited Linux kernel flaws.

The CNCF approved gVisor's application in late September, with Google announcing its donation plans shortly thereafter. The project's onboarding commenced last week, preceding Thursday's announcement of all three donations. Etienne Perot and Jing Chen, gVisor contributors and Google software engineers, sought to reassure the community about Google's continued commitment to the project in an October announcement.

We (the gVisor contributors at Google) have been pushing for this move internally with the expectation that this will accelerate gVisor's growth and adoption both within and outside of Google, in a similar manner as what has happened with Kubernetes.

Etienne Perot and Jing Chen, Google

Agent Substrate and OpenRL

Agent Substrate tackles a distinct problem: efficiently operating large fleets of autonomous agents without wasting resources during idle periods. The project introduces a runtime layer above Kubernetes that allows agents to be paused when inactive and reactivated when needed, eliminating the overhead of repeatedly provisioning and destroying Kubernetes Pods. It achieves this by decoupling an agent's persistent state from the computational resources running it, with gVisor serving as one secure isolation option.

Google released Agent Substrate 1.0 this week and made it generally available on Google Kubernetes Engine (GKE).

OpenRL concentrates on reinforcement learning, enabling developers to customize AI models using their own Kubernetes clusters. Launched in June, the project orchestrates the training and evaluation workflows that enhance model performance. It works alongside Agent Substrate, running training jobs inside isolated environments that can leverage gVisor for security.

Together, these three projects cover the execution, security, and training dimensions of autonomous AI agents.

Vendor-neutral governance and adoption barriers

The shift to CNCF governance will reshape gVisor's development and maintenance practices. Perot and Chen outlined plans in October to relocate the project's development and testing systems outside Google, grant external maintainers authority to merge contributions, and establish governance structures that prevent any single organization from wielding unilateral control.

In gVisor's CNCF application submitted via GitHub in September, Perot identified a gap in the project's user base. At one end sit major technology firms like Anthropic and OpenAI with resources to customize the technology; at the other are niche startups whose business models depend on secure code execution. The real opportunity, Perot argued, lies in reaching the much larger segment of companies in between.

Perot wrote in the application that such organizations "don't have the resources to invest specifically into security but would happily adopt an off-the-shelf, widely-available sandboxing runtime."

A related question concerns why more cloud providers have not adopted gVisor in their commercial services. In their October statement, Perot and Chen noted that Google, DigitalOcean, and Modal represent a small group of providers offering general-purpose compute backed by gVisor, despite evidence that customers are independently deploying it on other cloud platforms. They suggested that hesitation over Google's control of the project may account for this gap.

Perot and Chen wrote: "The remaining explanation of the lack of direct integration is likely the project's (pre-donation) governance risk."

Another significant hurdle involves the Linux kernel itself. Google and other organizations running gVisor maintain proprietary kernel modifications to boost performance, particularly for I/O-bound workloads. These enhancements are not universally deployed, potentially leaving other users with less efficient execution.

Upstream Linux maintainers declined to merge them while the project was owned by a single vendor.

Drew Bradstock, Google

Bradstock attributes the difficulty in getting kernel improvements accepted to gVisor's previous status as a Google-controlled project. Technical objections have also surfaced; LWN reported in 2022 that Linux kernel maintainers raised security and design concerns about a proposed performance enhancement, including doubts about its applicability beyond gVisor.

While such discussions do not consistently cite Google's ownership as the primary reason for rejection, moving to CNCF governance should help address concerns about single-vendor dominance.

For Google, this effort extends across all three projects. By transferring Agent Substrate and OpenRL to the CNCF alongside gVisor, the company aims to establish a unified platform for autonomous agents.

Vendor-neutral ownership lets cloud providers, hardware vendors, and independent software companies participate in the future growth of agentic workloads equally.

Drew Bradstock, Google