NitronEdge TechnologiesHelping Enterprises Harness Innovation
Security · Zenity · AI agents

Zenity secures agent sandboxes and feeds Microsoft risk

Zenity widened Microsoft agent coverage, launched sandbox security with NVIDIA OpenShell support and added a Zscaler route. What it means and what to do.

Zenity, the AI agent security company, made three announcements in three days around Microsoft Ignite. On 17 November 2026 it widened its coverage of AI agents across Microsoft products and began feeding its findings into Microsoft's own agent risk scoring through Sentinel. On 18 November it launched Zenity for Sandboxes, with day-one support for NVIDIA OpenShell. On 19 November it added a network option that inspects AI traffic passing through a secure web gateway, starting with Zscaler Internet Access.

What changed

Deeper Microsoft coverage. The list of Microsoft-hosted agents Zenity can watch now includes GitHub Copilot's coding agent on GitHub Enterprise, Copilot Studio, Agent Builder in Microsoft 365 Copilot, Copilot Cowork, Microsoft Foundry and Claude Enterprise inside Microsoft 365. Coverage is not the same everywhere. For Claude Enterprise and the GitHub coding agent, Zenity offers telemetry, findings and inline blocking. Copilot Studio agents get telemetry with limited blocking. Agent Builder and Cowork get telemetry and findings only. The Foundry integration has moved across to the new Foundry Agent Service.

Findings that change Microsoft's risk score. Zenity used to export raw logs to Sentinel. It now sends filtered alerts, and eligible findings feed the "dynamic agent risk" level Microsoft calculates for each agent. The first phase covers supported Copilot Studio and Foundry agents. Zenity says the same risk level will show up in Agent 365, Defender, Entra and Purview as Microsoft switches those views on, so treat that part as a roadmap rather than something you can use today. Ray Smith, who runs Agent 365 product at Microsoft, is quoted backing the integration.

Security for agent sandboxes. Coding and automation agents increasingly run inside sandboxes: short-lived machines where they can install packages, run commands and call the internet. Zenity for Sandboxes lists every sandbox and template, ties each one to its agent, owner, credentials, image and outbound network rules, and flags shadow sandboxes and agents running with no sandbox at all. It checks for real API keys left in images, wildcard egress, internet-reachable sandboxes and processes running as root. At run time it judges each action against what the agent was asked to do and whether untrusted input has crept into the session. If something looks wrong it can steer the agent, cut its network, freeze the sandbox for evidence, or kill it and rebuild from a clean image. Supported platforms include NVIDIA OpenShell, E2B, Daytona and Modal, plus the hosted agents from Anthropic, OpenAI and GitHub through the controls those vendors expose. There are starter templates for Claude Code, Codex and GitHub Copilot, and a new sandbox can begin in observe-only mode.

OpenShell integration. With NVIDIA OpenShell, the split is clear: OpenShell enforces, Zenity decides. The OpenShell gateway asks Zenity before creating a sandbox or changing its policy, and the supervisor checks with Zenity before traffic leaves. A baseline policy is attached to every new sandbox and attempts to strip it are refused.

A network route. For staff using public AI tools, including personal accounts, Zenity can now sit behind the company's web gateway as a chained proxy. It attributes each request to the corporate identity and can block prompts carrying secrets, source code or customer data before they reach the provider. No endpoint agent or new certificate is needed. If Zenity is unreachable, traffic carries on as normal. Only Zscaler Internet Access is supported today.

Why it matters

Most security teams can now list their human users and their servers. Very few can list their agents, and fewer still know which of those agents can run code. The sandbox launch goes after that gap. A coding agent with a production API key in its environment and open egress is a data leak waiting for a poisoned README file.

The Sentinel change matters for a different reason. It puts third-party runtime evidence into the risk score Microsoft itself shows to administrators. For a Microsoft-centred estate, that means one number per agent instead of two consoles that disagree.

The network option is the cheapest way to see shadow AI use, but it fails open by design. Treat it as visibility with some blocking, not as a hard control.

What this means for organisations in the Gulf, India and Africa

Many enterprises in the UAE, Saudi Arabia, India and South Africa have standardised on Microsoft 365 and are rolling out Copilot Studio agents built by business teams, often faster than security can review them. Sooner or later an auditor will ask the obvious question: who approved this agent, what can it touch, and where are the logs? The UAE PDPL, Saudi PDPL and India's DPDP Act all expect you to know where personal data goes. An agent that quietly pastes customer records into an outside model is a reportable problem, whatever the intent.

Two practical points. First, check where Zenity processes your telemetry before you connect regulated tenants, and ask the vendor in writing. Second, the network route only helps if your web traffic actually flows through Zscaler; many regional firms run other gateways or none at all for remote staff.

What to do next

  1. Build a list of every agent you run today: Copilot Studio, Foundry, GitHub Copilot, Claude, and anything your developers run in sandboxes. Name an owner for each.
  2. Look for sandboxes and build images that hold real credentials or allow any outbound traffic. Fix those first, whatever tool you use.
  3. If you are on Sentinel, ask Microsoft and Zenity which of your agents fall inside the initial dynamic risk integration, and test it on one Copilot Studio agent.
  4. Start any new agent policy in observe-only mode for two weeks, then turn on blocking for the actions that matter most.
  5. Decide whether you need endpoint, platform or network controls for public AI tools. Most firms need two of the three.

For a wider view of agent controls, see our Zenity partner page and our AI security services.

Sources

  1. Zenity Expands AI Agent Security Coverage Across the Microsoft Ecosystem (Zenity press release) (2026-11-17)
  2. Introducing Zenity for Sandboxes (Zenity blog) (2026-11-18)
  3. Zenity brings agent-aware security to NVIDIA OpenShell (Zenity blog) (2026-11-18)
  4. The network already sees every AI interaction (Zenity blog) (2026-11-19)
Related: Zenity
Enquire

Talk to us about Zenity

Get a free consultation from our AI, data and security specialists. Honest advice, clear options and a reply within 6 hours.

Reply timeOur team replies within 6 hours

Request a callback

Free consultation. Our team replies within 6 hours.

Your details are safe. Our team replies within 6 hours. By sending, you agree to our privacy notice.

WhatsApp