We're rolling out Claude company-wide through Entra SSO and want it to be the only approved AI service on work machines. A salesperson recently entered client names and complete worksheets into a personal ChatGPT account, so we're trying to reduce the risk of another data leak. We don't currently have DLP, Intune, or a large budget for new security tooling. We're a Fortinet shop with hybrid Active Directory, and I'm looking for practical interim controls using FortiGate web filtering, application control, DNS filtering, browser policies, and existing Microsoft or Entra capabilities. Policy and training are being handled separately; right now, I need the most effective technical controls we can deploy without major capital spending, along with any real-world advice about gaps or bypasses.
4 Answers
FortiGuard already has categories for AI chatbots, so you can start with a deny-by-default rule for those categories and explicitly allow the Claude domains. On FortiGate, also look at application control for DNS-over-HTTPS and use hostname inspection; otherwise browsers may resolve domains directly and bypass a DNS-based filter. It's a useful quick win, but test the exceptions carefully because vendor domains and features can change.
This needs layered controls rather than one blocklist. Use Entra Conditional Access so company data and services are accessible only from managed devices, and restrict personal-account sign-ins in managed browsers where possible. Lock down Claude connectors and grant them only the data permissions they need. Without DLP, you won't reliably stop someone from using a personal phone or home computer, so endpoint controls reduce the risk but cannot replace data classification, monitoring, and eventual DLP deployment.
The technical block is only part of the solution. Talk with employees about the tasks the approved tool doesn't support, because unmet needs are what often drive shadow IT. Then publish a clear acceptable-use rule, require acknowledgement, and define consequences for putting company or client data into unapproved services. Enforce it consistently with access and audit logs. Even with good filtering, policy, training, and accountability are what address the ways users can bypass network controls.
A proxy or DNS filtering service can do the same thing: block the AI category globally, then allow only the services you approve. Also remove local administrator rights, restrict application installation, and use enterprise browser policies to prevent unauthorized extensions. Endpoint scans through tools you already have, such as Defender or vulnerability-management software, can help find installed applications and browser add-ons that filtering won't catch.
Be sure to include AI features embedded in applications you already permit, such as browser assistants, meeting transcription tools, and CRM add-ons. Those may not appear as standalone AI websites and can require separate application or browser policies.