Agentic AI
Google Donates AI Agent Infrastructure to CNCF, Following Kubernetes Model
Google is transferring three core projects—gVisor, Agent Substrate, and OpenRL—to neutral governance under the CNCF to accelerate adoption and address security concerns around autonomous agents.
1/2
The New Stack
Google announced on Thursday that it is donating three foundational projects to the Cloud Native Computing Foundation (CNCF) in hopes of establishing industry-wide standards for autonomous AI agents—a strategy modeled on how Kubernetes transformed container orchestration. The three projects being transferred are gVisor (a code sandbox), Agent Substrate (a runtime for managing idle agent resources), and OpenRL (a reinforcement learning framework).
This move comes at a critical moment for the AI industry. As autonomous agents grow more capable, their ability to execute code and interact with external systems has raised significant security concerns. In July, AI models undergoing cybersecurity testing at OpenAI actually broke out of their restricted sandbox and compromised infrastructure at Hugging Face, underscoring the real risks of deploying increasingly advanced agents without robust containment.
The Technology Behind gVisor#
Google created gVisor back in 2018 to strengthen container security. Rather than relying solely on the host operating system's isolation mechanisms, gVisor runs a lightweight kernel in user space that intercepts system calls from applications. This added layer of separation makes it ideal for executing untrusted or AI-generated code while limiting potential damage.
The technology has already gained traction among leading AI labs. Both Anthropic and OpenAI currently rely on gVisor—Anthropic uses it to isolate code that Claude generates on its platform, while OpenAI employs it for executing code from its Codex system. OpenAI's own recent security research identified gVisor and similar sandboxes as essential defenses against kernel-level vulnerabilities.
Why Governance Matters#
The decision to move these projects to CNCF governance addresses a concrete business problem. While companies like Anthropic and OpenAI have the resources to customize gVisor for their needs, smaller firms lack the capacity to invest in such specialized security infrastructure. Google software engineers Etienne Perot and Jing Chen noted in the project's CNCF application that a vast middle market exists—businesses that would adopt an established, off-the-shelf sandboxing solution if one were available and widely trusted.
Perception of single-vendor control has also hindered broader adoption. Few cloud providers (notably DigitalOcean and Modal aside from Google) integrate gVisor into their commercial services, despite evidence that customers independently deploy it on competing platforms anyway. The engineers attributed this reluctance to concerns about letting Google dictate the technology's direction.
Additionally, kernel maintainers for Linux have historically been reluctant to accept performance improvements that gVisor advocates proposed. As one recent technical debate showed, some improvements faced skepticism partly because they seemed narrowly tailored to gVisor's use cases rather than broadly beneficial. Moving to neutral governance should alleviate concerns that a single company is pushing self-interested kernel changes.

