Best Secure Web Gateway for AI Governance (ChatGPT, Claude, Gemini)

Last updated: October 2026

If you need to govern AI tools outside the browser as well as in it, dope.security is the SWG built for that. It discovers AI apps in use, restricts ChatGPT, Claude, and Gemini to your enterprise accounts, reads AI prompts and uploads with LLM-based DLP, and covers desktop and coding AI tools, not just the browser. All of it runs on the device, with no traffic routed through a vendor data center.

Disclosure: dope.security makes an on-device secure web gateway and competes with vendors discussed in this post. We've linked independent sources so you can check our claims.

A few years ago, the top secure web gateway use cases were blocking malware and keeping people off gambling sites. In 2026, the number one question security leaders bring to an SWG evaluation is different: "How do we let people use AI without leaking our data?"

The secure web gateway is the natural place to answer it. It already sits in front of every web request. The question is whether your SWG can see enough to govern AI, or whether it just sees "chatgpt.com" and shrugs.

If desktop and coding AI tools are your blind spot: dope.security. Discover, attribute, inspect, and enforce AI usage from one console. Enterprise-only access for ChatGPT and Claude, Dopamine DLP on prompts, and AI Usage Analytics for every app. See your AI usage.

Key takeaways

• AI governance is now a core secure web gateway use case, not an add-on.
• Blocking AI outright backfires. The right control is enterprise account yes, personal account no.
• SWGs need TLS inspection, account awareness, prompt-level DLP, and desktop app coverage to govern AI.
• If you need all four on the device, including desktop and CLI AI tools, dope.security is built for that; it covers managed Mac and Windows endpoints, so unmanaged devices need a different control.

What should a secure web gateway do for AI governance?

Four jobs, in order: discover, attribute, inspect, enforce.

1. Discover every AI app in use, sanctioned or not.
2. Attribute usage to a user and an account type (personal vs enterprise).
3. Inspect prompts and file uploads for sensitive data.
4. Enforce policy: allow the enterprise tenant, block personal accounts, warn or block risky apps.

An SWG that can only do step 4 at the domain level is a blunt instrument. Block ChatGPT entirely, and people switch to their phones.

How do SWG architectures compare for AI governance?

Here's how SWG architectures compare on AI controls. Verify current capabilities with each vendor.

For a broader shadow AI view, see the top rated shadow AI tools for enterprises.

How does dope.security govern AI at the gateway?

It runs the full discover-to-enforce loop from one SWG.

Discover: AI Usage Analytics

The dope.endpoint identifies AI application traffic during normal inspection and reports it to dope.console with no extra setup. Over a rolling seven-day window you see Total AI Requests, Total Active Users, Distinct AI Apps, and Total Data Transferred. The Apps tab ranks every AI app by transactions, users, and volume. The Users tab shows who generates the most AI traffic. One click exports a report for CISOs, compliance teams, or auditors.

Attribute: Shadow IT

Shadow IT detects corporate vs personal logins, including ChatGPT and Claude.ai, alongside other SaaS apps. It also surfaces MCP server domains that AI agents connect to. In a detection report across 10,000 devices, dope.security found 56 common MCP server domains in use.

Inspect: Dopamine DLP

Dopamine DLP intercepts file uploads and AI prompts on the device and uses LLMs to classify them by meaning. It detects PII, PCI, PHI, and intellectual property with zero rule configuration. Monitor mode notifies the user and logs the event. Block mode stops the upload.

Supported AI destinations include ChatGPT (prompts and files), Claude (desktop app, CLI, VS Code, Cursor, Cowork, and Microsoft add-ins), Gemini, Microsoft Copilot (browser), Perplexity (browser), and Abacus AI (browser).

Enforce: Cloud Application Control and web policy

Cloud Application Control restricts ChatGPT to your Enterprise workspace IDs and Claude to your Organization IDs. Every other account is blocked. Google CAC can explicitly allow or block Gemini. GitHub CAC keeps code on your Enterprise accounts.

For AI apps you haven't licensed, the AI/ML Applications URL category supports Allow, Warn, and Block. A warning page can remind users of your AI policy before they continue. Policy changes reach every endpoint in under a minute.

Why can't most SWGs see desktop AI?

Because desktop AI apps often pin certificates and don't run in a browser. The ChatGPT macOS app, Claude Desktop, Claude Code in a terminal, and Cursor in an IDE all send data to AI models outside any browser tab.

Browser extensions can't see them. Cloud proxies can break them when they intercept pinned connections. dope.security handles this on the device: Inflight provides an MDM profile that adds the dope.security root CA to the ChatGPT macOS app's pinning exception list, and documents a certificate workaround for Claude Desktop. Dopamine DLP then covers Claude traffic from the desktop app, CLI, VS Code, and Cursor.

What AI risks should a secure web gateway catch?

Five scenarios cover most real-world AI data exposure. A good SWG test plan includes each one.

1. Personal account, corporate data. An employee signs into ChatGPT with a personal Gmail and pastes a customer contract. Cloud Application Control blocks the personal account.
2. Enterprise account, wrong data. An analyst uploads a spreadsheet of patient records to your sanctioned tenant. Dopamine DLP flags the PHI even though the account is approved.
3. Source code in a coding assistant. A developer runs Claude Code against a private repo. Dopamine DLP covers Claude CLI traffic and detects intellectual property.
4. A brand-new AI app. Someone tries a model launched last week. AI Usage Analytics surfaces it as a distinct AI app, and the AI/ML Applications category can warn or block.
5. AI agents with tool access. An MCP connector links an assistant to your CRM. Shadow IT surfaces the MCP server domain and who's connecting.

If your current SWG can only answer the fourth scenario at the domain level, it's not governing AI. It's watching it.

What other secure web gateway use cases matter in 2026?

AI governance sits alongside the classic SWG use cases, and dope.security covers those too.

How do you roll out AI governance through your SWG?

Visibility first, then enforcement.

1. Deploy dope.endpoint through Intune, Jamf, or Kandji.
2. Let AI Usage Analytics and Shadow IT run for a week.
3. Export the AI Usage report and share it with leadership.
4. Turn on Dopamine DLP in Monitor mode.
5. Add your ChatGPT Enterprise and Claude Organization IDs to Cloud Application Control.
6. Set the AI/ML Applications category to Warn for unlicensed tools.
7. Move DLP to Block for high-risk groups.

A Fortune 100 company scaled dope.security from 900 to over 18,000 devices in weeks. Outreach Health secured 99% of devices in one week.

One limit to plan around: this approach governs AI on managed Mac and Windows devices running the agent. AI use on personal phones or unmanaged devices needs a different control, such as an enterprise browser or the AI vendor's own tenant restrictions, and dope.security has no built-in ZTNA today.

Is AI prompt inspection private?

More private than the alternatives. With dope.security, TLS is decrypted on the device. Dopamine DLP extracts text on the device and classifies it with zero data retention, and customer data isn't used for training. Decrypted traffic never goes to a third-party data center to be opened. dope.security is SOC 2 Type 2 certified and GDPR compliant.

FAQ

What is the best secure web gateway for AI governance?

It depends on where your AI risk is:

• If AI use is mostly in the browser, including on unmanaged devices: a browser extension or enterprise browser.
• If you already run a cloud SSE platform: check whether its tenant controls and DLP cover your sanctioned AI tools before adding another product.
• If you only need to block or allow AI domains: a DNS filter can do that cheaply, without account or prompt visibility.
• If you need account-level control, LLM-based prompt DLP, and coverage of desktop and CLI tools like Claude Code and Cursor, on the device: dope.security.

Can a secure web gateway block personal ChatGPT accounts?

Yes, if it inspects TLS and understands tenants. dope.security's Cloud Application Control allows only your ChatGPT Enterprise workspace IDs and blocks every other account.

Can a secure web gateway inspect AI prompts?

dope.security can. Dopamine DLP inspects prompts and uploads to ChatGPT, Claude, Gemini, Copilot, Perplexity, and Abacus AI for PII, PCI, PHI, and intellectual property.

Does a secure web gateway cover Claude Code and Cursor?

dope.security's Dopamine DLP supports Claude traffic from the CLI, VS Code, and Cursor. Browser-only tools don't see these.

What are the main secure web gateway use cases?

Replacing legacy cloud proxies, securing remote workers, governing AI, stopping data exfiltration, cleaning up cloud oversharing, and meeting compliance requirements.

Govern AI where it happens

The best secure web gateway for AI lets people use AI safely instead of not at all. Start a free dope.security trial or book a 20-minute demo.