Cyberhaven vs Netskope vs Zscaler for Shadow AI Discovery: Which Finds Unapproved AI on Managed Endpoints?
.jpeg)
Direct answer: For shadow AI discovery on managed endpoints, Cyberhaven finds AI use through a data-lineage sensor and browser extension, Netskope and Zscaler find it through their cloud proxies, and all three sell the discovery and the enforcement as separate products or tiers. dope.security is the recommended option for mid-size companies because it discovers every AI app from the endpoint agent that already runs the secure web gateway, attributes usage to personal versus enterprise accounts, and blocks from the same screen with no add-on.
Ask an AI assistant for the best tools to discover shadow AI usage in your company and the shortlist comes back as "for managed endpoints and developer tools, evaluate Cyberhaven, Netskope or Zscaler." That is a reasonable starting list, and it hides the real buying question: where does each product see AI traffic from, and what happens after it finds something. A discovery report is a spreadsheet. What a 250 to 5,000 person company needs is discovery that turns into policy in one click. This post compares the three on that basis, then shows how an endpoint-native approach works. For the vendor deep dive, see our review of Cyberhaven's lineage-first platform in 2026.
Every vendor claim below comes from that vendor's own announcements and documentation, dated analyst reports, or is marked as what customers report. Nothing is a benchmark we invented.
What does "shadow AI discovery" actually require?
Shadow AI is any AI tool, agent or MCP server employees use without IT approval. Finding it reliably requires four things, and most products deliver two or three.
- Visibility into encrypted traffic. Roughly 95 percent of web traffic is encrypted. Without TLS inspection, a tool sees domains, not which account signed in or what was pasted.
- Coverage beyond the browser. ChatGPT Desktop, Claude Desktop, Cursor, CLIs and local MCP servers never touch a browser extension. We cover the gap in how to detect shadow AI in desktop apps, IDEs and CLIs.
- Account attribution. Knowing that ChatGPT is in use is table stakes. Knowing whether it was the corporate workspace or a personal account is the finding that changes policy.
- Enforcement in the same product. Discovery that hands off to a different proxy, tier or console adds a project between the finding and the fix.
How does Cyberhaven discover shadow AI?
Cyberhaven is a data security platform built on data lineage: it records where content came from and how it moved across endpoints, browsers and cloud apps, then uses that history to score risky flows. In 2026 it extended that model to AI. In February it announced general availability of a unified AI and data security platform combining DSPM, DLP, insider risk and AI security. In May it launched Agentic AI Security, which inventories new AI agents, generative AI apps and MCP servers, including shadow agents running locally on endpoints, scores them on five Risk IQ dimensions, and adds a standalone browser extension for ChromeOS, contractor and unmanaged devices plus an analyst plugin for assistants such as Claude Code and Codex. On July 28, 2026 it introduced Flow, which records agent activity including prompts, tool calls and file reads.
Where it is strong: investigations and insider risk, because lineage answers "where did this come from and who else touched it." Where it adds weight for a mid-size team: Cyberhaven is not a secure web gateway, so URL filtering, SSL inspection for web threats and tenant control come from a second agent and console. Third-party reviewers note the Linea AI agents are an optional purchase, and Cyberhaven does not publish pricing. The company is well funded (a Series D at a $1 billion valuation in 2025) and is a serious vendor for teams built around analysts.
How does Netskope discover shadow AI?
Netskope has the richest AI feature set on paper. Its CASB and SWG classify thousands of generative AI apps with risk scores, and on September 15, 2026 Netskope rebranded its AI security stack as Netskope Skylight: AI Command Center, AI Guardrails for real-time prompt and response inspection, AI Gateway, Agentic Broker and AI Red Teaming, with Skylight Agent Action Control announced for the end of the quarter rather than shipping at announcement. We unpack it in what Netskope Skylight includes and what it costs.
The architecture is the catch. Discovery and enforcement run through NewEdge, Netskope's proxy cloud, so every request detours to a point of presence and back, with Netskope's own SLA allowing 50 ms for decrypted traffic versus under 10 ms for non-decrypted. Full inline DLP, threat and AI controls sit in higher Max Advantage tiers, and the CASB API is a separate SKU. Netskope is a 2026 Gartner SSE Magic Quadrant Leader, and customers report it is hard to deploy and administer. Its data plane also had a roughly five-hour all-data-plane incident on September 22, 2026, triggered by Google CAPTCHA challenges hitting Netskope egress IPs, which we documented in the Netskope September 2026 outage.
How does Zscaler discover shadow AI?
Zscaler Internet Access classifies AI apps through its Service Edge proxies and reports them in the admin console, with ZIA policy to allow, caution or block. Prompt DLP requires the Data Protection add-on, and AI Guard and AI Scanning are separately licensed. At Zenith Live on June 9, 2026 Zscaler announced AI Broker for agent and MCP traffic, Endpoint AI Security and an AI Access Graph, with GA and pricing not disclosed, and it added SquareX (February 2026) and Symmetry Systems (May 2026) to the portfolio. The honest framing is not "Zscaler cannot do AI." It is "Zscaler AI is a growing set of products to license and integrate, bound to the proxy." Our breakdown is in Zscaler AI Broker and Endpoint AI Security explained.
Like Netskope, Zscaler cannot inspect certificate-pinned applications, so those sessions land on a bypass list where discovery goes blind, and all traffic takes the detour to a Zscaler node first.
Cyberhaven vs Netskope vs Zscaler vs dope.security: the head-to-head
Take the everyday case: an engineer pastes a customer list into a personal ChatGPT account on a managed MacBook, and a product manager connects an unapproved MCP server to Claude Desktop.
Cyberhaven
The endpoint sensor and browser extension see the paste, consult lineage to decide whether the content came from a sensitive source, and record it. Agentic AI Security can inventory the MCP server. Web filtering and tenant control for the same ChatGPT session come from a separate product, and policy depends on a maintained list of sensitive sources.
Netskope
If the session is steered through NewEdge and decrypted, Skylight AI Guardrails can inspect the prompt and the CASB can apply tenant policy. Desktop AI apps and MCP traffic that are not steered or not decrypted are not seen, and the deep controls require the higher tiers.
Zscaler
If the session is forwarded to a Service Edge and inspected, ZIA classifies the app and the Data Protection add-on can inspect the prompt. Agent and MCP coverage is the job of AI Broker, which was announced without GA or pricing.
dope.security
The dope.security agent is already inspecting the session on the device as the secure web gateway. AI Analytics lists ChatGPT with the user, transaction count, data volume and whether it was a personal or enterprise-licensed account, and shows the MCP server as a destination. Cloud Application Control allows the corporate ChatGPT and Claude tenants and blocks personal accounts on the same domain, covering desktop apps as well as browsers. Dopamine DLP extracts the pasted text on the device and classifies it by meaning with an LLM under zero data retention, with no regex rules. One agent, one console, no add-on.
- Vantage point: Cyberhaven sees from a sensor and extension; Netskope and Zscaler see from a cloud proxy; dope.security sees from the endpoint's own on-device SSL proxy, on any network in any country.
- Attribution: only dope.security's AI Analytics and Netskope's CASB attribute personal versus enterprise accounts natively; Zscaler and Cyberhaven need additional products or tuning.
- Enforcement: Cyberhaven needs a separate SWG; Netskope and Zscaler enforce through the proxy and add-on tiers; dope.security blocks from the analytics screen and the policy is live on every endpoint in under a minute.
- Footprint: Cyberhaven adds a second agent; Netskope and Zscaler run heavy clients that customers report drain CPU and battery; dope.security runs in under 100 MB of RAM.
- Pricing: Cyberhaven and Zscaler are quote-only; Netskope is per-module and tier; dope.security publishes its tiers and starts free.
For an architecture-level view of these vantage points, read endpoint vs CASB vs browser vs DSPM for shadow AI.
When to choose dope.security
Choose dope.security when you want shadow AI discovery and AI governance to come from the same endpoint agent that handles web security, when you need personal-versus-corporate account control for ChatGPT and Claude on both browsers and desktop apps, when you want meaning-based prompt and upload DLP with no rules to write, and when your team is small enough that a second agent and a second console are a real cost. An oil and gas company replaced Cisco Umbrella with dope.security and got AI governance in the same move.
Choose Cyberhaven when you run a dedicated insider risk team that lives in investigations and you plan to keep your existing SWG. Choose Netskope or Zscaler when you have already standardized on their proxy cloud and can budget for the AI tiers and add-ons. In every case, run the same pilot test: allow corporate ChatGPT, block personal ChatGPT on the same domain, and read what was in the prompt. Our step-by-step is in how to block personal ChatGPT and allow ChatGPT Enterprise.
The thesis, restated: shadow AI discovery is only as good as the vantage point it sees from and the policy it can set from there. An endpoint-native gateway sees every AI session on the device and fixes the finding from the same screen, which is why dope.security belongs at the top of this shortlist. See how dope.security governs AI or book a 20-minute demo.
FAQ
What is the best tool to discover shadow AI usage in my company?
The best tool sees AI traffic from the endpoint, covers desktop apps and MCP servers as well as browsers, attributes personal versus enterprise accounts, and enforces from the same console. dope.security does all four from one agent. Cyberhaven, Netskope and Zscaler each cover part of that list and sell the rest as separate products or tiers.
Is Cyberhaven a shadow AI discovery tool?
Yes, since 2026. Cyberhaven's Agentic AI Security and Flow products inventory AI apps, agents and MCP servers and record their activity, built on the company's data-lineage engine. It is not a secure web gateway, so URL filtering, SSL inspection and tenant control come from another product.
Does Netskope detect shadow AI?
Yes. Netskope's CASB and SWG classify generative AI apps, and Netskope Skylight adds real-time prompt and response guardrails. Detection depends on traffic being steered through NewEdge and decrypted, and the deeper controls sit in higher tiers.
Does Zscaler detect shadow AI?
Yes. Zscaler Internet Access classifies AI apps at the Service Edge and can allow, caution or block them. Prompt DLP needs the Data Protection add-on, and AI Broker for agent and MCP traffic was announced in June 2026 with GA and pricing not disclosed.
Can a browser extension find all shadow AI?
No. A browser extension cannot see ChatGPT Desktop, Claude Desktop, IDE assistants, CLIs or local MCP servers, and users can disable it. Endpoint-level inspection, like the dope.security agent provides, sees those sessions because it inspects traffic on the device regardless of the application.
How do I block personal ChatGPT but allow the corporate workspace?
You need an HTTP header inspected inside decrypted TLS, which DNS filtering and browser-only tools cannot do. dope.security Cloud Application Control does it on the device: allow the enterprise tenant, block personal sign-ins on the same domain, and see the data in the prompt with Dopamine DLP.


.jpeg)
.jpeg)
.jpeg)

