In addition to the question of how smart we can make our AI, engineers are starting to ask how can we keep our agents from doing stupid (or even worse, malicious) things with our keys? Google have tackled this challenge by taking a very open approach to keeping everyone in. Today (Nov 5, 2026), Google Cloud it is donating Agent Substrate and gVisor as part of their Sandboxed projects to the Cloud Native Computing Foundation. This was announced at KubeCon North America in Salt Lake City.
Problem We Cannot Avoid
AI agents are not chatbots. They are coding agents with the ability to call tools, control browsers and interact with databases. They are described as running untrusted code and tools and need to be isolated from their host and network. An agent can easily be fooled by a poisoned website or document and execute malicious code. There is no way of avoiding this with careful prompts – only using an effective architecture.
Traditional containers share a kernel with the host which is fine if the code running is trusted. However, if the code was created a second ago by a statistical model parsing the internet then it would be irresponsible to share the kernel.
From Simple To Sophisticated
This is the latest step in Google’s roadmap. Last year they launched Agent Sandbox, an open source Kubernetes primitive designed for AI agents. It provides isolation for each task in the form of a sandbox with kernel level security. It was based on gVisor with support for Kata Containers. In August this year Google Cloud and Anyscale released an experimental gVisor sandboxing library for Ray clusters. As always, Google are continuing to develop the infrastructure and are choosing to give it away. With Agent Substrate, Google are providing the most sophisticated solution yet. They describe it as an orchestrator for AI agents with less than 500ms latency for resuming sessions. Agent Substrate has been released as a 1.0 product. This is important as if an agent is too slow due to isolation, the first impatient developer will turn it off! During the conference there was some comment about how sandboxing should provide a secure default, whilst freeing up unused compute.
Why Donate?
Whilst cynics may point out it is a clever move from Google to donate infrastructure to a neutral foundation, they are not wrong.
Trust: Enterprises will not allow their sensitive workflows to run within proprietary black boxes. If the code is open source then it can be audited.
Standardisation: If Agent Substrate and gVisor become standard for isolating agents on Kubernetes, Google are setting the standard for the whole ecosystem.
Best Practice: Even with open standards, they will perform best when they are run by their creators. Agent Sandbox can be used directly in GKE.
Neutral Governance: Giving the project to the CNCF shows competitors and customers that no single vendor can pull the carpet.
Sandboxed means that this is the entry level project within the foundation. This means it is promising but not fully mature.
A Note of Scepticism
The headline statements are taken directly from Google’s post and there is no mention of any independent testing. Whilst 500ms+ latency to resume sessions is impressive, real world scenarios can be quite complex.
Sandboxing will not solve everything. It can limit the damage caused by a compromised agent but it won’t prevent them from doing something malicious with the permissions they have been given. Even if the agent is fully contained it can still send the incorrect file via email. Containment is required but should be combined with least privilege credentials, audit logs and manual checks.
Conclusion
The main thing this confirms is that the industry is moving forward. The first stage of the AI boom saw whoever could create the most capable model win. The next stage will see whoever can make these models safe enough to deploy at scale win. Whilst infrastructure, containment and orchestration aren’t exciting, they are essential.
Google have decided to push themselves into this space by being open and asking the community to help them build their walls. This is the right move to make. Security should always be transparent.
Ultimately, the agents are coming. But by making their solution open source, they are ensuring they are coming within nice, well built, auditable containers. Well done Google, but let’s make sure they actually are!
