78% of CISOs have already yanked permissions from an AI agent running in production. That number comes from a survey of 300 security executives, and it frames everything about Mirantis’s timing. This week at KubeCon North America 2026 in Salt Lake City, Mirantis launched Agent Platform GA — a Helm-installable governance layer designed to give organizations actual control over the AI agents running on their Kubernetes clusters before something goes wrong.
The Governance Gap Is Real
Kubernetes was designed for containers and services, not for autonomous agents with LLM API access that act on your infrastructure at 3 AM. The result is a governance gap that is not theoretical anymore. According to NeuralTrust’s State of AI Agent Security 2026, 88% of enterprises reported AI agent security incidents in the past 12 months. Only 17.5% of organizations can automatically stop a misbehaving agent in production. The industry moved fast on AI agents in ops. It did not move fast on guardrails.
CNCF’s own data puts 49% of cloud-native teams running AI agents in production — and KubeCon NA 2026 made it clear the ecosystem knows it has a governance problem. The event added a dedicated “AI Inference + Agentic” track this year for the first time. Multiple sessions covered the same uncomfortable question: Kubernetes cannot tell which AI agent took action, and that matters when agents have kubectl access.
What Agent Platform Actually Does
Mirantis Agent Platform is a self-hosted platform, installed via a single Helm chart into any conformant Kubernetes cluster. The control plane, sandboxes, database, and audit trail all stay in your infrastructure — no data leaves. It ships with four governance primitives that have been missing from most AI agent deployments:
- Sandboxing: Each agent runs in an isolated environment. Policy defines exactly what it can reach — specific namespaces, tools, APIs. Access outside that scope is blocked at the network level, not just logged afterward.
- Spend controls: Token budgets are configurable at the organization, project, and individual agent level. When a budget ceiling is hit, the agent stops. Autonomous loops do not run up unbounded LLM API bills.
- Audit trails: Every agent action is logged — not the final outcome, the full action history. This is what makes post-incident investigation possible instead of guesswork.
- Kill switch: One click cuts network access for every agent in the cluster simultaneously. No hunting through individual sandbox configs during an incident.
Human Approval Is the Feature, Not the Limitation
Mirantis ships a built-in agent called Prism that handles infrastructure operations. The workflow: Prism detects a production issue, diagnoses the root cause, and proposes a specific remediation. A human reviews and approves — or rejects — the proposed fix. Only after approval does Prism execute anything.
This is not autonomous SRE. That is the point. The agent does the investigation — scanning cluster state, correlating logs, identifying the probable cause — and the human makes the execution decision. It compresses the time from “alert fired” to “root cause understood” dramatically, while keeping a person in control of what actually changes in production. That distinction matters when you are talking about Kubernetes clusters running payment services or healthcare data pipelines.
Agent Platform is not limited to Prism. Organizations already running third-party agents — custom Claude agents, Cursor Agent SDK deployments, or homegrown tooling — can route them through the same governance layer. A single dashboard covers every agent in the fleet regardless of origin.
Deployment and Availability
Installation is a single Helm chart. Mirantis supports any conformant Kubernetes distribution — MKE, EKS, GKE, AKS, OpenShift. A 30-day evaluation is available; enterprise pricing requires contacting Mirantis directly. The company is pairing the platform with Forward Deployed Engineering for teams that want senior Mirantis engineers to help design and operationalize their agent workflows before handover.
The Bigger Picture at KubeCon
Mirantis is not alone in recognizing this category. At KubeCon this week, Google donated Agent Substrate and gVisor to the CNCF Sandbox specifically for sandboxing untrusted agents. Solo.io’s agentgateway handles MCP-based agent traffic governance. kagent provides a Kubernetes-native agent lifecycle layer. These are complementary primitives, not direct competitors — which suggests the cloud-native community is building governance infrastructure in layers, the same way it built security and observability in layers.
Moving fast on autonomous agents without a governance layer is not bold engineering — it is how you end up in that 78% statistic. Human-in-the-loop is not a crutch for teams that are not ready for AI ops. It is the responsible architecture for running agents on infrastructure that matters. Full announcement details are available on the Mirantis press release.
