Everyone is building AI agents today. An agent that handles support tickets, writes and deploys code, or pulls reports from your CRM can be stood up in a few days. Getting it into production, however, often takes months. Not because the models aren't good enough, but because nobody knows exactly how to keep such an agent under control.
Rubrik, known to most of our customers for backup, cyber recovery and identity resilience, has an answer: Rubrik Agent Cloud.
In this article we explain what it is, why it matters for Belgian organizations, and why it makes perfect sense that a data security vendor is the one taking this step.
The problem: agents are LLMs with tools
An AI agent is essentially a large language model with access to tools: databases, APIs, file systems, SaaS applications. That combination is powerful, but also risky. Agents operate at machine speed, frequently use non-human identities with broad access to enterprise data, and are by definition non-deterministic. They can hallucinate, make mistakes, or find a workaround when something is blocked.
Rubrik itself cites a well-known example: a coding agent that accidentally dropped a production database during a code freeze. Dev Rishi, General Manager of AI at Rubrik, also described an agent that noticed its MCP connector had been disabled, opened a browser window on its own and reached the data via drive.google.com anyway. Classic, static rules are no match for that.
The result is that in many organizations AI governance exists mainly on paper. Every new agent goes through a committee, every exception is reviewed manually, and by the time the green light arrives the agent is already obsolete. The figures Rubrik quotes speak for themselves:
What is Rubrik Agent Cloud?
Rubrik Agent Cloud is an enterprise control layer that sits between your applications, your agents and the underlying LLMs. It was announced at Rubrik FORWARD in October 2025 and has been generally available since February 2026.
The promise boils down to three verbs: monitor, govern, remediate. Or, in Rubrik's own words: 'Unleash Agents. Not Risk.'
What sets the platform apart from most AI security tools is the foundation it is built on. Rubrik Security Cloud has managed the intersection of data, identity and, more recently, through the Predibase acquisition, models for years.
Agent Cloud enriches agent telemetry from sources such as OpenAI, Amazon Bedrock, Microsoft Copilot Studio and Anthropic Claude with exactly that context: which data does this agent touch, under which identity, and what does a healthy state of that data look like?
The four pillars
- Agent Observability: Automatically discovers every agent and MCP server in your environment, regardless of framework. Shows when they were created, how heavily they are used, their risk profile and which tools and data they touch – all with a full audit trail.
- Agent Identity: Determines who can do what with which agent. Scopes MCP and tool access per user and group, federates with Okta and Microsoft Entra ID, and mints a short-lived, minimally scoped token for every tool call. No more standing permissions.
- Agent Runtime Security: Powered by SAGE: natural-language policies are enforced in real time on both the inputs (prompts) and outputs (responses and tool calls) of your agents. Out-of-the-box guardrails such as PII detection are available immediately.
- Agent Rewind: When an agent does break something, Rubrik correlates that action with a healthy snapshot so you can reverse the damage in minutes: a dropped database, deleted Salesforce opportunities or a corrupted codebase.
SAGE: governance in natural language
At the heart of the platform is SAGE, the Semantic AI Governance Engine Rubrik unveiled at RSAC 2026. The idea is as simple as it is far-reaching: instead of writing deterministic rules that cannot cope with the richness of natural language, you simply write your policy in plain English. Small, specialised language models then interpret the intent, parameters and likely impact of every agent action and check them against that policy.
Concretely, every MCP tool call passes through three checkpoints before it executes:
- Behavior: SAGE semantically analyses what the agent is trying to do and whether that violates policy.
- Access: the MCP Gateway enforces access policy at the infrastructure layer, enriched with SAGE's intent and risk context.
- Identity: the agent session is authenticated and a short-lived token is minted that is valid for this single tool call only.
A write or modification without an explicit policy scope is blocked before it ever happens. And if something still goes wrong despite all that, there is Agent Rewind.
Agent Rewind: the undo button only Rubrik has
This is where Rubrik's DNA makes the difference. There are by now several observability and governance tools for agents. But no other vendor can actually reverse a destructive agent action, because doing so requires immutable, verified snapshots of your data – and Rubrik already has those.
Agent Rewind correlates an erroneous action with a healthy snapshot in Rubrik Security Cloud and restores exactly what is needed. Not a full restore of an environment, but the dropped table, the deleted Salesforce records or the overwritten files. For customers already running Rubrik today, this means their existing backup investment suddenly protects their AI agents too.
New since June 2026: Agent Cloud for Claude Code and Claude Cowork
At Rubrik FORWARD 2026, Rubrik extended the platform to coding agents, specifically Anthropic's Claude Code and Claude Cowork. These agents write, push and deploy code autonomously, while classic DevSecOps controls assume there is always a human in the loop. Rogue commits, repository ransomware, prompt injection and IP exfiltration thereby become possible at a scale version control alone cannot protect against.
Rubrik adds the following for this use case:
- Agent Rewind for the codebase itself, including immutable recovery when a bad action outruns what Git can track.
- Backup, monitoring and restoration of the agentic harness: CLAUDE.md files, system prompts and tool configurations that govern how an agent behaves. Configuration drift is detected and tied to a healthy snapshot.
- Integration with the Claude Compliance API to monitor how Claude is being used across the organization.
On top of that, in mid-September 2026 Rubrik launched Rubrik Code Guardian, a further step towards protecting source code and developer workflows.
What this means for Belgian customers
Important in practice: Agent Cloud is not tied to a single AI stack. Rubrik supports three integration paths:
- Custom: builds via a standalone AI gateway for agents built on frameworks such as LangChain or with direct LLM calls to OpenAI, Claude or Bedrock.
- Low-code platforms: backend integrations with Microsoft Copilot Studio among others, announced at Ignite 2025. New Copilot agents are automatically discovered, monitored and protected, including their access to OneDrive and SharePoint.
- Endpoints: discovery and governance for agents running directly on workstations.
For the many Belgian organizations currently experimenting with Copilot Studio agents on top of Microsoft 365, that second path is the most tangible. It is precisely there that shadow AI emerges fastest: a business user builds an agent, gives it access to a SharePoint site, and IT only finds out when something goes wrong.
Why this matters to your organization
At Agisko we see the same pattern among our customers that Rubrik describes. There is no shortage of AI pilots; the step to production stalls on governance questions nobody has a conclusive answer to.
Who is accountable if an agent deletes something?
How do you prove to an auditor, or under NIS2, what an agent did?
How do you recover quickly without rolling back the entire environment?
Rubrik Agent Cloud answers those questions not with another layer of policy on paper, but with infrastructure that enforces the policy at the moment an agent acts. And crucially, it builds on a platform you may already have.
If you use Rubrik Security Cloud for backup and cyber recovery today, the foundation for Agent Rewind is already in place.
Our recommendation? Start with visibility. The Agent Observability pillar maps out which agents are already running in your environment. It's often more than you think. From there you determine where governance and identity controls are needed most, and which agents touch critical data for which Agent Rewind is a must.
Want to know more?
Agisko is a Rubrik partner and has been helping Belgian organizations with Rubrik Security Cloud for years, from architecture and implementation to managed services.
Want to know how Rubrik Agent Cloud fits into your AI roadmap, or curious which agents are already active in your environment today? Get in touch with our team for a demo or an exploratory conversation.
