Home / Guides / block-employees-pasting-pii-into-chatgpt

How to stop employees from pasting PII into ChatGPT

Network DLP misses browser LLM paste. Roll out on-device Browser Shield, workspace policy, and telemetry so customer data never leaves the device unmasked.

Published 2026-08-28 · NoeticGuard engineering notes

Security teams hear the same story: an analyst pastes a customer export, ticket thread, or spreadsheet into ChatGPT to "summarize faster." Acceptable-use policies help culture; they do not intercept the keystroke. Classic DLP was built for SMTP and sanctioned SaaS proxies — not employee browser sessions on chatgpt.com.

Why blocking the domain fails

  • Productivity pressure — teams route around blanket blocks with personal accounts or alternative UIs.
  • TLS end-to-end — proxies often cannot inspect prompt bodies without brittle SSL break-and-inspect.
  • Catalog lag — new models (Claude, Gemini, DeepSeek) appear faster than CASB URL lists update.

Rollout checklist (IT / SecOps)

  1. Publish AI acceptable use — approved tools, no customer PII in public LLMs, incident reporting. Template: employee AI policy guide.
  2. Deploy Browser Shield — Chrome extension runs @noeticguard/core on-device. See Browser Shield docs.
  3. Force-install via MDM — Chrome Enterprise policy for managed browsers. Steps: MDM & Chrome Enterprise.
  4. Paste workspace policy key — each extension uses an ng_pub_ key from Dashboard → API Keys for config sync and quota telemetry only.
  5. Enable Media Guard (Startup+) — block screenshots and file uploads on supported LLM hosts when paste masking is not enough. Media Guard.
  6. Measure, don't guess — dashboard quota and Startup+ audit logs show masked-event volume per workspace.

What Shield does not do

Browser Shield masks text before LLM egress in supported web UIs. It is not a replacement for vaulting secrets in your product backend — developers should still use the SDK for API-integrated assistants. For the DLP architecture gap, read why classic DLP fails on ChatGPT.

Measuring success

  • Masked-event volume on the workspace dashboard (quota meter on Free; audit logs on Startup+).
  • Fewer security tickets about “customer data pasted into ChatGPT.”
  • Policy acknowledgment rate after publishing the AI acceptable-use template.

Related comparisons

Evaluating cloud DLP, network appliances, or other LLM egress approaches? See our comparison hub for local-first masking vs cloud DLP and other approaches.

Get startedRead the docs