Enterprises Block Unauthorized ChatGPT, Gemini, Claude Access

Kaspersky published a June 10, 2026 guide on detecting and disabling "shadow AI", employees using unsanctioned AI assistants such as ChatGPT, Claude, Gemini, and meeting bots, which it ties to data-exfiltration and compliance risk. The guide is structured as a practical control checklist rather than a threat report, listing vendor domains and technical indicators alongside the controls that catch them.

Kaspersky's recommendations operate at two layers. At the network layer: block the "AI Services" category in the NGFW or web filter, and reroute DNS queries for specific domains, for example claude.ai, anthropic.com, *.anthropic.com, and api.anthropic.com, and the equivalent OpenAI domains, away from production endpoints. At the endpoint: enforce browser extension allowlists through policy, and use EDR/EPP or application control to scan for installed desktop clients such as claude.exe or ChatGPT.exe and to inspect browser history. The guide explicitly notes that a hard ban tends to backfire, driving users to personal devices, third-party wrappers, or proxy services, and recommends standing up an approved, auditable AI option before turning on enforcement.

Commercial products converge on a similar architecture. Per its product page, TeamPrompt pairs a browser extension that scans prompts for sensitive data with network-level blocking via Cloudflare Gateway, and advertises deployment through Google Admin, Microsoft Intune/JAMF, or GPO, with Monitor, Restrict, and Full Lockdown modes. The vendor also claims 40-plus detection rules and coverage of roughly 18 AI tools; those figures are vendor-stated and unverified here.

The hard part is not blocking, it is blocking without breaking work. Broad "AI Services" category enforcement is the fastest control to deploy but generates the most false positives and the strongest incentive to route around IT. The durable pattern in Kaspersky's guidance is to instrument first, run prompt-DLP and category controls in monitor mode to measure real usage and false-positive rates, then enforce, and to pair any block with a sanctioned alternative so the underlying demand has somewhere legitimate to go.

• •Adoption of browser-extension prompt DLP versus network-only blocking as the default shadow-AI control.
• •False-positive rates when enforcing broad AI-category blocks, and their effect on user workarounds.
• •Whether enterprises pair enforcement with approved internal AI tools, the mitigation Kaspersky flags as decisive.