A shadow AI discovery tool that only watches the browser or the network will miss the coding assistants, CLI agents and local MCP servers your engineers run, where the riskiest AI use now sits. Start with the free inventory already sitting in your OAuth and sign-in logs. Then buy a product with an endpoint agent. For most companies with a real developer population that means Harmonic Security, or Netskope's endpoint discovery if you already pay Netskope. Nudge Security is the right buy for a SaaS-heavy company with few developers, and it is the only one here that lists a price on its own website. A browser extension such as Akamai Workforce Protector (formerly LayerX) is a good enforcement layer and a poor sole inventory.
Shadow AI discovery is the job of finding every AI app, assistant and agent employees use, sanctioned or not, and who uses it. Prices below were checked on each vendor's live page on November 17, 2026.
| Product | Detection method | Sees IDE and CLI tools | Sees MCP servers | Published price (checked Nov 17, 2026) | Pick it if | Skip it if |
|---|---|---|---|---|---|---|
| Harmonic Security | Browser plus endpoint agent | Yes, on the Guide plan | Yes, on the Command plan | None; demo only | You have developers and want prompt-level coaching | You need a number before a first call |
| Netskope One AI Command Center | SSE inline plus endpoint and server discovery | Yes, with endpoint discovery on | Yes | None; contact sales | You already route traffic through Netskope | You have no SSE and want a small tool |
| Nudge Security | Mailbox, OAuth and identity, browser extension | No | Not listed | $5/user/month, 150 to 1,500 users | Your AI risk is SaaS sign-ups and OAuth grants | Your risk is on developer laptops |
| Akamai Workforce Protector (LayerX) | Browser extension | No (PWA desktop apps only) | Not listed | $8.50/user/month on AWS Marketplace, 50-user minimum | You want in-browser redaction and blocking | It would be your only inventory |
| Microsoft Defender for Cloud Apps | Network and endpoint log discovery | Partly, via Defender for Endpoint | Not listed | No list price on Microsoft's pricing page | You already license it and run Defender for Endpoint in active mode | Your fleet runs a third-party antivirus |
Which Detection Method Sees Which Kind of AI Use?
Each of the four methods sees a different slice, and none sees everything. This matrix is the reason the verdict above favours an endpoint agent.
| Surface | Browser extension | CASB or SSE (inline) | Endpoint agent | Identity or OAuth discovery |
|---|---|---|---|---|
| Web chat (ChatGPT, Claude, Gemini in a tab) | Yes, including prompt text | Yes, if traffic is steered and decrypted | Partly; sees the browser, not always the tab | Only if the user signs in with a work identity |
| Desktop AI apps | No, except PWA apps | Yes, if steered and decrypted | Yes | Only with SSO or OAuth sign-in |
| IDE assistants (Cursor, GitHub Copilot) | No | Partly; sees the API domain | Yes | Partly; sees the account, not usage |
| CLI agents (Claude Code, Codex CLI) | No | Partly; sees the API domain | Yes | Partly, through API keys or SSO |
| Local MCP servers | No | No for local stdio servers; remote ones show as traffic | Yes | No |
| Third-party apps holding OAuth grants | No | No | No | Yes |
Two rows explain most bad purchases. A local MCP server runs as a process on the developer's laptop and talks to its client over standard input and output, so no proxy ever sees it. And an OAuth grant that a meeting-notes bot got eight months ago generates no browser or proxy event today, yet it still reads your mail. Only the identity method finds the second, and only an endpoint agent finds the first.
What Can You Inventory for Free Before Buying Anything?
You can build a usable first inventory in about two weeks from systems you already pay for. Do this before any vendor demo, because it tells you which column of the matrix your risk actually sits in.
Start with OAuth grants. Google Workspace's OAuth log events record the application name, the OAuth client ID, the scopes granted and the user who granted them. Filter for AI vendors and for scopes that read mail, files or calendars. In Microsoft 365, the Entra enterprise applications list does the same job, and Microsoft's own security-for-AI guide shows how to filter it to agent identities built in Copilot Studio and Foundry.
Next, export sign-in logs. Entra keeps sign-in and audit logs for seven days on the free tier and 30 days on P1 or P2, and upgrading does not bring back expired data. Route them to storage this week, or the history you need for a baseline disappears.
Then pull expense and card data: every reimbursement and corporate card line for OpenAI, Anthropic, Cursor, Perplexity, Midjourney and the other vendors your engineers name. A personal ChatGPT Plus subscription expensed every month is a shadow AI user who has told you who they are.
Last, check your existing proxy or CASB catalog. If you already have Microsoft Defender for Cloud Apps, its catalog has a Generative AI category with more than a thousand AI apps and 90+ risk factors. Turn on the filter and look at daily users.
The free inventory has the same blind spot as the identity column: it cannot see a CLI agent running on an API key bought with a personal card, or a local MCP server. That gap is what you are paying a vendor to close.
What Each Product Actually Covers
Harmonic Security: the strongest endpoint coverage, no public price
Harmonic is the pick for a company with developers, because it is the only product here whose published plans name IDEs, desktop apps and MCP discovery as separate tiers. Its pricing page lists three plans and no prices. Explore gives prompt-level visibility across 1,000+ applications, covering web apps, APIs, OpenTelemetry, desktop apps and AI agents. Guide adds endpoint coverage for desktop apps, IDEs and copilots, with inline coaching that warns, nudges or blocks in under 200 milliseconds (a vendor claim). Command adds MCP client and server discovery, read, write and execute permissions for agents, and a local MCP gateway.
Read the fine print on unmanaged devices: the same page describes its agentless Claude controls as "visibility only" for mobile and unmanaged devices. If a large share of your AI use happens on personal laptops, Harmonic will tell you it is happening and will not stop it.
Who should not buy it: a company that needs a budget number before a first sales call, and a company with almost no developers, which would be paying for Guide and Command features it never uses.
Netskope One AI Command Center: the right answer if you already own Netskope
If your traffic already runs through Netskope, turn on its endpoint AI discovery before you buy anything else. The AI Command Center launch in June 2026 extended the Netskope One Client to scan installed applications, running processes and listening ports on managed endpoints, finding AI agents, local models and browser extensions. A separate eBPF agent intercepts TLS-encrypted AI traffic on corporate VMs and Kubernetes nodes, which covers server-side agents your platform team runs.
Netskope's documentation says it detects MCP servers from app events and transaction log events, and adds a third source when endpoint discovery is on. That puts MCP servers in the same inventory as SaaS AI apps. Netskope publishes no price for AI Command Center; you will need a quote.
Who should not buy it: a company with no SSE today. Buying a full SSE platform to get an AI inventory is a network re-architecture project, with a timeline measured in quarters.
Nudge Security: the cheapest honest start, blind on laptops
Nudge finds AI through identity: it reads users' mailboxes in Google Workspace or Microsoft 365 for sign-up and billing emails, maps OAuth grants, API tokens and service accounts, and adds a browser extension for login data, per its pricing page and how-it-works page. Because it reads historical mail, it finds tools people signed up for months ago and forgot.
It is also the only vendor here that publishes a price on its own website. Essential is $750 a month flat for up to 150 users. Growth is $5 per user per month for 150 to 1,500 users, billed annually. Above 1,500 users it is a custom enterprise agreement. AI agent discovery and governance is listed as included, but a separate "AI agents" feature is marked as a paid add-on with no listed price.
Who should not buy it: a company whose AI risk sits on developer laptops. Mailbox and OAuth discovery see the Cursor subscription; they do not see what Cursor sends, which local MCP servers it launches, or a Claude Code session running on a personal API key.
Akamai Workforce Protector (formerly LayerX): enforcement in the browser
Akamai closed its acquisition of LayerX on July 2, 2026, after announcing it in May at about $205 million. The product is now sold as Akamai Workforce Protector, an extension for Chrome, Edge, Firefox and Safari that also covers the agentic browsers ChatGPT Atlas, Perplexity Comet and Dia, and PWA desktop apps such as Teams, Slack and VS Code. The product page does not mention MCP.
Akamai's own page lists no price. LayerX's AWS Marketplace listing is $8.50 per user per month with a 50-user minimum, cancellable any time, with no refunds.
The extension is good at the thing a browser is good at: seeing the prompt before it is sent and redacting or blocking it. For web chat it gives you more detail than any proxy. That is why it is a strong second purchase. As your only inventory, it is the loser in this comparison, because the surfaces it cannot see (native IDEs, CLI agents, MCP) are the ones growing fastest. An Akamai ownership change is also a reason to put a continuity clause in the contract before you sign.
Who should not buy it: anyone expecting it to inventory developer tooling.
Microsoft Defender for Cloud Apps: free if you have it, fragile in practice
If you already license Defender for Cloud Apps, use it for the first pass. It discovers generative AI apps, scores them, and lets you unsanction an app so Defender for Endpoint blocks it on onboarded devices. Purview's DSPM for AI adds prompt-level insight for supported AI sites, but only for devices onboarded to Purview.
The catch is the endpoint dependency. One admin on Microsoft's Q&A forum saw no generative AI usage at all on any endpoint, while Defender for Endpoint ran in passive mode alongside a third-party antivirus, and the reply asked them to test with Defender for Endpoint in active mode. If your EDR is not Microsoft, check this before you trust a quiet dashboard. Microsoft does not list a standalone price on its security pricing page; check what your agreement already includes.
Who should not rely on it: a company running another vendor's EDR, or a Mac-heavy engineering team.
Identity-first alternatives such as Reco and Obsidian Security cover similar ground to Nudge from the SaaS posture side, and Zenity focuses on agents built inside low-code platforms. Evaluate them against the same matrix.
What Does It Cost at 1,000 Employees?
Take one workload: 1,000 employees, 150 of them developers, Microsoft 365, a mixed Windows and Mac fleet, and an annual contract. At list prices checked November 17, 2026:
| Product | Annual cost at 1,000 users | Covers developer laptops |
|---|---|---|
| Nudge Security (Growth) | $60,000 ($5 x 1,000 x 12) | No |
| Akamai Workforce Protector (AWS Marketplace) | $102,000 ($8.50 x 1,000 x 12) | No |
| Harmonic Security | Quote only | Yes (Guide or Command) |
| Netskope AI Command Center | Quote only | Yes, with endpoint discovery |
| Microsoft Defender for Cloud Apps | Depends on your existing licence | Partly |
The two products with public prices are the two that cannot see the developer laptop. Ask Harmonic and Netskope for a quote at exactly this workload, and ask whether endpoint coverage can be licensed for the 150 developers only. That split is often the cheapest way to close the gap.
Why You Cannot Trust the Rankings
Nearly every "best shadow AI detection tools" list on the first page of results is written by a company in the category. Nudge publishes one, Reco publishes one, and dope.security publishes a 12-platform comparison. Research for this guide found no independent head-to-head test that ran these products against the same fleet.
Run your own. Give each shortlisted vendor the same 20 test cases on a pilot group of laptops: a personal ChatGPT login in Chrome, the Claude desktop app, Cursor with a personal API key, Claude Code from a terminal, a local MCP server that reads a file share, a Gemini side panel, a meeting bot with a calendar OAuth grant. Count what each tool reports within 24 hours.
How Do You Get Prompt Capture Past Legal and the Works Council?
Plan for prompt capture to need a works council agreement in Germany and a DPIA in the UK and EU, and design the deployment so it can pass both. The browser and endpoint products above can capture prompt text, which is employee monitoring.
The useful German precedent cuts toward this. In January 2024 the Hamburg Labour Court (24 BVGa 1/24) held that a works council had no co-determination right over ChatGPT guidelines, partly because browser-based AI used without installed software made monitoring impossible and the company had no access to the AI provider's data. A discovery extension or endpoint agent that logs prompts gives the company exactly the access the court found missing, which is what Section 87(1) No. 6 of the Works Constitution Act covers. The ruling came in summary proceedings, so take this to employment counsel before you deploy.
The UK regulator's guidance on monitoring workers requires a DPIA for high-risk monitoring and recommends one even when the risk is not high. Workers must be told about monitoring before it starts, and an employer must pick the least intrusive means that achieves the aim.
A design that tends to clear both reviews:
- Collect metadata by default: app, user, timestamp and data classification match, with no prompt body.
- Capture prompt text only on a policy match, for example when a DLP rule fires on customer data or source code.
- Pseudonymise user identity in dashboards, with re-identification limited to a named investigation role.
- Set a short retention period for captured prompts and write it into the agreement.
- Tell employees first, in plain language, with the list of what is and is not collected.
Ask each vendor which of those five settings their product supports out of the box. A product that cannot run in metadata-only mode will be hard to deploy anywhere with a works council.
What Do You Do With the Inventory Once You Have It?
Sort every discovered tool into one of three treatments, and coach by default. Blocking first tends to push employees onto personal phones, where no tool on this list can see them.
Coach when the tool is fine but the data is not, with a nudge that says the company has a ChatGPT Enterprise tenant and links to it. Redact when the tool is sanctioned and the user is about to paste a customer list or a secret. Block a tool only when it has terms you cannot accept (training on inputs with no opt-out, no data residency you can use) or a known security problem, such as the browser extension flaw covered in our ThisBash analysis.
For MCP servers, the inventory feeds an allowlist enforced at the client, the approach set out in our MCP governance buyer's guide.
Measure whether it is working with four numbers, baselined over the 30 days before you turn on coaching:
- Weekly active users of unsanctioned AI tools, counted per user, not per event.
- The share of AI traffic that lands on sanctioned enterprise tenants.
- Repeat triggers per user after a first coaching nudge.
- Count of OAuth grants to AI apps with mail, file or calendar scopes.
If unsanctioned weekly users are not falling within 90 days, the sanctioned tool is the problem. People route around a coaching nudge when the approved alternative is worse.
How to Decide Without Regret
The criteria that predict regret in this category:
- If the free inventory shows most AI spend coming from developers, an identity or browser tool will disappoint you. Buy endpoint coverage.
- An existing Netskope or Microsoft E5 deployment changes the answer. Use what you have before adding an agent to every laptop.
- Every endpoint and extension product here needs a managed device to enforce anything. Contractors on their own laptops show up only through identity.
- In a works council jurisdiction, a product that cannot run metadata-only is a product you cannot deploy.
- LayerX changed hands in July 2026. Put a change-of-control clause in every contract in this category.
This Week: Export Entra or Google Workspace sign-in and OAuth logs to storage, and pull 12 months of expense lines for AI vendors.
This Month: Build the first inventory from those sources, then write the 20-case pilot test and send it to Harmonic and, if you run Netskope, to your Netskope account team.
Before Q1 Close: Get the works council or DPIA review done on a metadata-first design, then run the pilot on 50 laptops that include your developers.
The Bottom Line
This category is following the path CASB took a decade ago: first an inventory of cloud apps, then controls, then consolidation into larger platforms. The difference is that the riskiest AI use has moved off the browser and onto the developer's machine, into tools that act on their own. Buy for that surface, and run your own test before you trust anyone's ranking, including this one.
Start with the OAuth log export; it is free and it expires.
Continue Reading
- Personal Claude Accounts Expose Work Laptops to ThisBash
- MCP Server Governance: Enforce at the Client, Not the Registry
- Claude Duet checks Gmail, Slack and HubSpot every few hours
- Adobe Took Rilo's Team. Its Slack Token Never Expires.
- Your Claude Enterprise DLP Hooks May Let Unchecked Prompts Through
- MAS Makes Financial Firms Inventory AI Across Their SaaS Tools
- Agent Authorization: Standing Privilege Is the Whole Problem
