At Zenith Live in June 2026, Zscaler announced AI Broker for agent and MCP traffic, Endpoint AI Security, and an AI Access Graph. It was a serious set of announcements from the SSE category leader. It was also a set of announcements with GA timing and pricing not disclosed, routed through the same cloud a user's traffic already detours to. The thesis of this post is simple and testable: Zscaler governs AI and agent traffic by adding another SKU that runs through its cloud, while dope.security governs ChatGPT, Claude, Gemini, and MCP traffic natively on the device, with no extra license and no backhaul.

The short answer: Zscaler's AI Broker and Endpoint AI Security are real products on a real roadmap, and Zscaler is a 2026 Gartner SSE Magic Quadrant Leader. But they are new, separately licensed capabilities bolted onto a proxy architecture that forwards every request to a Zscaler service node first. dope.security does AI governance as a foundation, not an add-on, and inspects on the endpoint so there is no detour and no second subscription.

What Zscaler actually announced

Credit where it is due. Zscaler is a category leader, its shadow-AI discovery is strong, and the Zenith Live lineup was ambitious. AI Broker is positioned to govern traffic from AI agents and Model Context Protocol servers. Endpoint AI Security extends controls toward the device. AI Access Graph maps how AI is used across the environment.

Here is the part that matters for a buyer. At announcement, GA timing and pricing for these products were not disclosed. In Zscaler's own AI governance stack, prompt DLP already runs through the Data Protection add-on, and AI Guard and AI Scanning are separately licensed. So the honest framing is not "Zscaler has no AI story." It clearly does. The framing is "Zscaler's AI story is more products to license, integrate, and route through the cloud." That is a different buying decision than turning on a capability you already own.

The architecture tax: it still routes through a service node

Zscaler's model forwards traffic to a ZEN or Service Edge node for inspection. That is the foundation the new AI products sit on. As modules stack, customers report latency climbing, and cert-pinned apps like Microsoft 365, WebEx, and Dropbox are documented as hard to inspect through the proxy. An AI Broker that routes agent and MCP traffic through the cloud inherits that same detour: device to point of presence to destination and back, on every request.

For AI workloads this is not a rounding error. Agent traffic is chatty. MCP calls chain. Every hop you add to a model request, a tool call, or a retrieval step shows up in how fast the thing feels. If the inspection point is a cloud node far from the user, the AI tax is the model's own latency plus the detour.

dope.security inspects on the device and flies direct, with no point-of-presence hop on the path of a prompt or an agent call. Want to see the detour in numbers? See how Fly Direct works or book a 20-minute demo and measure the gap on your own connection.

The fastest AI governance is the kind that does not add a round trip to every agent call.

The licensing picture: AI governance as more SKUs

Zscaler's pricing has been moving toward Z-Flex consumption licensing, which the company highlighted in its Q4 FY26 results in September 2026 with more than 770 million dollars in total contract value tied to it in the quarter. Consumption licensing can be flexible, and it can also make it harder to predict what AI governance will cost once AI Broker, Endpoint AI Security, Data Protection, AI Guard, and AI Scanning are all in the bill of materials.

Layer on the acquisition pace. Zscaler bought SquareX in February 2026 and Symmetry Systems for 175 million dollars in May 2026. Acquisitions add capability, and they add integration surface: more consoles, more agents, and more stitching between parts that were built by different teams. The AI governance a buyer wants should feel like one product, not a portfolio to assemble.

Agent and MCP traffic: what an "AI Broker" has to do

Governing agent and MCP traffic is a real problem and Zscaler is right to chase it. An AI agent connects to tools and MCP servers, pulls data, and acts. The controls you want are discovery of which agents and servers are in use, the ability to allow sanctioned connections and block unsanctioned ones, and inspection of the data moving through those connections.

The question is where that control lives. A cloud broker can only govern the agent and MCP traffic that is actually routed through it. The MCP server a developer wires up locally on their laptop, the desktop AI client, the CLI that talks straight out: those are exactly the connections that bypass a cloud chokepoint. Watching the endpoint is how you see them. For the full picture of why the device is the right vantage point, see how on-device visibility surfaces every AI app.

Zscaler's cloud-routed AI governance vs dope.security on the device

Both approaches want the same outcome: see AI and agent usage, allow the sanctioned, block the risky, and stop sensitive data from leaving. They differ on where the work happens.

dope.security is the clear pick when the goal is AI governance that is native, fast, and priced as one thing. Zscaler is a strong proxy that is adding AI products on top of the proxy. Those are different products.

How dope.security governs AI and MCP on the device

dope.security runs a lightweight agent, under 100 MB of RAM, that does SSL inspection, web filtering, Cloud Application Control, and AI-powered DLP locally. For AI and agent traffic that means three layers working together.

Discovery. AI Analytics surfaces every AI app in use, sanctioned or not, with per-user transactions and volume and a personal-versus-enterprise account breakdown, over a rolling window, with no proxy logs to grep.

Policy and tenant control. From the same screen you can block an app or enforce enterprise-only access, and Cloud Application Control locks SaaS and AI tools to your corporate tenants. This is the control a three-layer AI governance model is built around: discover, apply SWG policy, then restrict the tenant.

Inspection. Dopamine DLP reads prompts and attachments on the device and blocks a paste or upload that contains regulated data, with a plain-language summary of what it caught. It pairs with sanctioned-versus-unsanctioned SaaS governance so the enterprise tools keep working while the personal ones get stopped.

A Fortune 100 company scaled dope.security from 900 to over 18,000 devices in weeks, deploying silently through Intune with no manual configuration. That is the deployment profile of a foundation, not an add-on you stage over several quarters.

The short version

Zscaler's AI Broker and Endpoint AI Security are credible announcements from a category leader, and anyone evaluating them should take them seriously. Just read them for what they are: new, separately licensed capabilities that route through the same cloud your traffic already detours to, with GA and pricing still to be nailed down. dope.security takes the opposite path. AI governance, tenant control, and DLP run on the device as one product, so you see the agent and MCP traffic a cloud broker misses, you pay for one subscription instead of a stack of AI SKUs, and nothing takes a detour to a data center and back.

Frequently Asked Questions

What is Zscaler AI Broker?

Zscaler AI Broker is a product announced at Zenith Live in June 2026 to govern traffic from AI agents and Model Context Protocol servers. At announcement, its GA timing and pricing were not disclosed. It is part of a broader Zscaler AI lineup that also includes Endpoint AI Security and AI Access Graph. Because it sits on Zscaler's cloud-proxy architecture, it governs the agent and MCP traffic routed through Zscaler, not connections that bypass the cloud, such as locally wired MCP servers on a laptop.

Does Zscaler charge extra for AI governance?

Zscaler's AI governance is spread across separately licensed pieces. Prompt DLP runs through the Data Protection add-on, AI Guard and AI Scanning are licensed separately, and AI Broker and Endpoint AI Security had undisclosed pricing at announcement. Zscaler is also shifting toward Z-Flex consumption licensing. dope.security includes three-layer AI governance and Dopamine DLP in one subscription with no per-capability add-on.

Can a cloud AI broker see MCP servers running on a laptop?

Only if that traffic is routed through the cloud. An MCP server a developer connects locally, a desktop AI client, or a CLI that talks straight out can bypass a cloud chokepoint. Seeing those connections requires watching the endpoint. dope.security inspects on the device, so it surfaces agent, desktop, and locally-connected MCP activity that a cloud-only broker would miss.

Is dope.security faster than Zscaler for AI traffic?

dope.security inspects on the device and adds no network detour, so a prompt or an agent call does not take a hop to a point of presence and back. Zscaler forwards traffic to a service node, and customers report latency rising as modules stack. For chatty agent and MCP workloads, removing the detour is where the on-device model shows up most. You can measure the gap on your own connection with the Fly Direct speed test.

Is Zscaler a bad product?

No. Zscaler is a 2026 Gartner SSE Magic Quadrant Leader with strong shadow-AI discovery, and its AI announcements are credible. The trade-off is architectural. Zscaler adds AI governance as separately licensed products on top of a cloud proxy that forwards your traffic, while dope.security runs AI governance and DLP natively on the endpoint as one product. Which fits depends on whether you want AI controls as an add-on stack or as a foundation.

How does dope.security handle AI governance differently?

dope.security treats AI governance as a foundation, not a bolt-on. One lightweight agent discovers every AI app with a personal-versus-enterprise breakdown, Cloud Application Control restricts tools to your corporate tenants, and Dopamine DLP inspects prompts and uploads on the device. It runs under one console on one subscription, with no backhaul and no separate AI SKUs to assemble.

See how dope.security governs AI on the device or book a 20-minute demo and compare it against what a cloud broker can actually see.