KubeCon: Google Hands Over Two Isolation Tools to CNCF

At KubeCon + CloudNativeCon 2026, Google entrusted gVisor and Agent Substrate to the Cloud Native Computing Foundation (CNCF) as part of its sandbox program. Held in Salt Lake City, the conference also showcased tools to manage agent permissions, their deployments, and expenditures. These announcements come after reported test environment breaches by OpenAI and Anthropic.
OpenAI Reports a Breach at Hugging Face
In July, OpenAI stated that models undergoing cybersecurity evaluations had breached the boundaries of their test environment and accessed Hugging Face's production systems. Nine days later, Anthropic reported three incidents involving Claude: from misconfigured environments, the models had connected to the Internet and reached the systems of three external organizations.
Anthropic later acknowledged that its protection relied solely on the configuration of the test environment. According to Omdia, leaks caused by tool calls represent the most frequently mentioned security threat, at 51%. The firm also indicates that 32% of organizations have a sandbox in production, meaning an isolated environment or a code execution layer.
Red Hat Blocks Unauthorized Connections
Agents generate code and then execute it, often before any human verification. Their isolation imposes a reverse trade-off compared to what was favored for a decade: containers, followed by serverless functions, had increased the number of hosted workloads at the cost of weaker separation than that of virtual machines. For security, platform, and infrastructure teams, treating this code as untrusted requires protection superior to that of a standard container, even with fewer workloads per server.
A user-space kernel, a micro-virtual machine, or a full virtual machine can provide this separation. Virtual machines thus find a use with Kubernetes that goes beyond legacy applications. With gVisor, a distinct kernel operates as a user-space program: all requests from the code to the operating system pass through it before reaching the server's kernel. Agent Substrate manages these environments at scale; Google intends this set for executing generated code without exposing servers hosting other workloads.
The CNCF also admitted OpenShell, Nvidia's open execution environment, into its sandbox program. This environment governs the use of tools, inference requests, and workload execution through deterministic controls. Its level of protection depends on the rules written by platform teams.
Jessica Forrester, senior distinguished engineer at Red Hat, presented its integration into the company's AI platform. In Red Hat OpenShift AI 3.5, OpenShell initially prohibits any outgoing connections and then grants access only to explicitly authorized destinations. The rules are stored in version-controlled YAML files and enforced by the Linux kernel, rather than by the agent. According to Red Hat, this setup limits the data that a compromised agent can extract.
Companies Must Identify Each Agent
The majority of agents used in enterprises remain installed on employees' computers, often outside of IT visibility. Omdia notes that 44% of organizations have employees using open frameworks locally, such as LangChain or CrewAI, without IT oversight. According to the firm, 52% of organizations fear leaks from unauthorized agent tools; half also worry about compliance breaches related to their autonomous actions.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Platform and security teams are called to take charge of these usages, whether they anticipated them or not. The responsibility for actions remains tied to those delegating tasks: they must be able to observe and restrict the activity conducted in their name. Abdel Sghiouar from Google Cloud dedicated a talk to the issue of agent identity in clusters, while responses mostly come from commercial offerings.
Octopus separately registers humans and agents to maintain certain validations under exclusive human responsibility. At Akuity, the commercial platform Argo CD allows the agent to borrow the identity of its user, but with reduced rights. Dapr represents a response from the CNCF: this graduated project, primarily maintained by Diagrid, allows since version 1.18, released in June, to cryptographically sign each step of an agent's process. It uses SPIFFE, an open standard for workload identity.
Nirmata Avoids Validation by Deployment
Jim Bugwadia, CEO of Nirmata, describes an evolution from simply formulating instructions to implementing governance mechanisms, now aiming for agents that are active 24/7. He deems it incompatible with scaling to require authorization at each deployment. Launched in July by Nirmata, AIControls applies Kyverno policies to agents, a CNCF engine already used for Kubernetes: developers can deploy without individual validation if they adhere to IT and security limits.
Broadcom imposes another restriction in the foundations of Tanzu Platform agents: access is denied by default in isolated environments. Only Model Context Protocol (MCP) models and servers, which connect to tools and data, managed by IT, are accessible. Developers thus have agents without access to unapproved models or sources.
For testing, Flox produces a Nix environment based on the developer's request that precisely sets the versions of libraries and tools. The code tested locally by an agent finds the same software in the cluster, which, according to Flox, reduces the time lost due to production failures caused by environmental differences. LocalStack, for its part, emulates cloud services, including AWS, without real accounts, real data, or production credentials. The company presents this approach as a way to limit damage to real resources and avoid cloud testing costs.
Nutanix Attributes Consumption to Teams
Financial tracking was also among the expressed needs, with measurement by agent or task. Nutanix Agent Gateway makes visible the teams and agents consuming tokens. According to Nutanix, the gateway offers faster approved access to new models than circumventing IT, reduces unauthorized usage, and ties expenses to the relevant teams.
User adoption is particularly visible on GitHub. OpenClaw has accumulated around 395,000 stars, more than Linux in 15 years; less than a year after its launch, it belongs to the two most starred software projects, excluding tutorials and link collections. Harness, from DeepSeek, does not reach this top group but surpassed in 15 days what Linux or React gathered over several years. These tools are aimed directly at end users.
Organizations are eager to adopt agents to enhance productivity and operational capacity. Poorly controlled adoption can lead to difficulties and unfavorable media coverage; conversely, a competitor that integrates them into its products and development can gain a commercial lead that is hard to bridge. However, the devices presented mainly concern individual positions, execution in clusters, and model gateways. Among the mentioned offerings, observability providers are the only ones accessing production, solely for telemetry read.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.