Commvault and Zscaler Team Up to Govern and Recover Enterprise AI Agents

Commvault and Zscaler announced a strategic partnership on November 10, 2026, to help enterprises govern production AI agents and roll back harmful or unintended activity. The deal combines AI-use discovery and policy enforcement on Zscaler's side with recovery for data, applications and identities on Commvault's. The companies say they want to close the operational gap between controlling an agent and repairing what it changes.

The stakes are straightforward. Experimental agents are becoming production systems with direct access to business data. When that happens, security and IT teams face a set of questions the industry hasn't fully answered yet: which AI tools do employees and developers actually use, what data do those tools share, which activities should be permitted or blocked, and — the hardest one — can unwanted changes be reversed quickly when something goes wrong?

What does each company bring?

Zscaler Security for AI identifies enterprise use of AI applications and enforces policies for both human users and AI agents. Commvault's AI and cyber-resilience technology restores data, applications and identities after attacks, disruptions or outages. Stitched together, the two companies propose a path from detection through recovery — a full cycle for agent incidents rather than a halfway measure.

That division of labor is the partnership's immediate significance. Zscaler would provide visibility and enforcement. Commvault would address restoration after an incident. The connective tissue is the idea that preventive controls need a recovery path for the changes that inevitably slip past them.

What are the planned use cases?

The companies named three collaboration areas: AI discovery, runtime governance, and agent detection and recovery. The announcement presents these as planned workstreams rather than a generally available joint product. That caveat matters for buyers evaluating timelines.

The three workstreams break down as follows:

• Enterprise-wide discovery. One workstream would expand discovery and maintain continuous visibility into active AI applications, models and tools across the enterprise. You can't govern what you can't see, and shadow AI use remains a widespread blind spot.
• Guardrails against unsanctioned activity. A second workstream would create guardrails that flag unsanctioned AI activity and prevent sensitive-data exposure before it spreads.
• Real-time protection plus recovery. The third pairs real-time protection for autonomous agent operations with rapid, clean recovery following a security event or a risky AI action.

Each workstream targets a distinct failure mode: unknown tools, known-but-dangerous behavior, and damage already done. Whether all three ship as integrated joint capabilities remains to be demonstrated.

What is AgentRecover?

Alongside the partnership, Commvault gave a more detailed view of its recovery approach in a separate AgentRecover announcement. The new capabilities support Amazon Bedrock AgentCore and Microsoft Foundry — two platforms that provide services for building, deploying and operating production agents. AgentRecover is designed to audit agent activity, restore agents and recover affected data and systems across environments.

Commvault Chief Technology and AI Officer Pranay Ahlawat framed the problem in blunt terms. Agents do not behave deterministically, he said, and a single agent decision can trigger changes across data, applications and infrastructure on multiple platforms. Such an unintended action can propagate at machine speed, affecting data integrity, application availability and business operations.

That machine-speed propagation is what separates agent risk from ordinary software risk. A human making a bad change works at human pace. An autonomous agent making the same mistake can touch dozens of systems before anyone notices, which is why auditing and recovery have moved from nice-to-have to core requirements.

How does the recovery actually work?

For supported platforms, AgentRecover is designed to discover agents and map the data, applications, identities, configurations and infrastructure with which they interact. That dependency map is intended to reveal protection gaps — the connections nobody documented when the agent was first deployed.

Its auditing functions serve a second purpose. They are designed to establish what an agent did and what it may have affected across platforms, helping teams prioritize recovery work when time is short. Knowing where the blast radius ends is half the battle in an incident.

Recovery is designed to cover both the agent itself and the affected resources, concentrating on what the agent changed rather than restoring everything indiscriminately. Customers can also use AgentRecover alongside Commvault's identity recovery and change auditing to examine agents' actions, the data they changed and the identities through which they acted. That identity dimension matters: agents acting through service accounts or credentials leave an identity trail that ties the whole investigation together.

When will it be available?

Commvault is targeting early access for Microsoft Foundry in November 2026 and for Amazon Bedrock AgentCore in December 2026. Those dates apply specifically to AgentRecover's rollout. The Zscaler announcement does not make them availability dates for the partnership's planned work, so joint capabilities could arrive on a different schedule entirely.

Why this matters now

The partnership connects preventive controls with a recovery path for the changes that slip past them — a pairing the enterprise AI market has largely lacked. Most agent governance conversations stop at detection and blocking. Few vendors have committed to answering the uglier question: what happens after an agent has already corrupted data or disrupted an application in production?

The two-company structure also reflects where the responsibility sits today. Security platforms like Zscaler own visibility and enforcement at the access layer. Recovery and resilience vendors like Commvault own the cleanup. Enterprises that adopt AI agents at scale will likely need both halves, and the market signal here is that vendors are starting to build the connective tissue themselves.

The central execution test, as the announcement itself implies, is turning planned use cases into working discovery, governance and agent-recovery capabilities for joint customers. Roadmap partnerships in security have a mixed track record, and this one spans two vendors with distinct product stacks, separate sales motions and different platform integrations. Early-access dates for AgentRecover — November 2026 for Microsoft Foundry and December 2026 for Amazon Bedrock AgentCore — will offer the first concrete evidence of whether Commvault's recovery half is ready, with the joint Zscaler work to follow on its own timeline.