We're rolling out Claude across the company through Entra SSO and want it to be the only approved generative AI service on work machines. We recently discovered an employee had pasted client names and complete worksheets into a personal ChatGPT account, so preventing another data leak is now a priority.
We don't currently have DLP, Intune, or DNS filtering, and budget constraints make a full DLP project unlikely in the short term. We're a Fortinet shop with hybrid Active Directory, so I'm looking for practical interim controls using FortiGate web filtering, application control, browser policies, or similar tools. Policy and training are being handled separately; right now I'm mainly interested in technical measures that don't require major new spending. What has worked for others?
5 Answers
Lock down the endpoints as well. Remove local administrator rights, restrict application installation, and use enterprise browser policies to prevent unapproved extensions and personal account sign-ins. Scan workstations with whatever Defender or vulnerability-management tools you already have to find existing extensions and applications.
Use a default-deny approach: block AI services at the FortiGate or proxy, then explicitly allow the approved service. Custom domain indicators can help catch popular tools. Review the block list regularly because new services and AI features inside existing products won't always appear under obvious AI domains.
FortiGuard already has an AI chatbot category, so you can start by blocking that category and adding exceptions for the approved Claude domains. Also enable application control for DNS-over-HTTPS and inspect hostnames, otherwise browsers may resolve sites directly and bypass a DNS-based block. Treat this as a best-effort control rather than a complete solution.
Blocking sites alone won't solve the whole problem. Talk with users about what the approved tool does and doesn't support, provide a way to request legitimate exceptions, and get management and HR to back a policy with meaningful consequences. That combination usually reduces shadow-tool use more effectively than a firewall rule by itself, while giving you time to implement DLP later.
The technical controls need to be paired with access restrictions and a clear policy. Use Entra conditional access so company resources are accessed only from managed devices, and limit permissions for Claude connectors to the minimum required. Otherwise someone can simply move the data to a personal computer, phone, home browser, or another allowed application.
That makes sense. I'm pushing for DLP and stronger access controls, but approval and funding are outside my control, so I'm trying to build a reasonable interim setup with what we already have.