Use Workplace AI Safely: What You Must Never Paste
Purpose
AI assistants can help outline, summarize, rewrite, brainstorm, and explain. They can also expose information if you enter data that the service, your organization, or the task is not approved to process. Treat every prompt as a disclosure to an outside system unless your organization has explicitly approved the tool and the specific data type. The UK National Cyber Security Centre notes that prompts may be visible to the provider and may be used in service development, depending on the product and settings. Approval for one AI tool does not automatically cover another tool, a personal account, a browser extension, or a new feature.
Warning signs
Stop before submitting if your prompt contains names linked to private details, internal-only material, copied email threads, unpublished business plans, or text you would not post publicly. Also stop if the tool asks you to paste authentication information, upload a real customer record, disable a security control, or move work into a personal account. Unexpected requests for urgent access or verification may be phishing. CISA recommends resisting urgency, checking the sender or site through a known route, and reporting suspicious messages.
Safe steps
- Confirm that the AI tool is approved for workplace use. Use the company-provided account and normal sign-in page. If you are unsure, ask your manager or support team before entering work information.
- Classify the information before prompting. Never enter passwords, MFA codes, API keys, access tokens, encryption keys, customer data, health data, financial data, source code, contracts, confidential documents, legal advice, personnel records, or other restricted information.
- Reduce the prompt to the minimum needed. Replace real people, organizations, account numbers, dates, and unique events with invented placeholders such as Customer A or Project X. Removing a name alone is not enough if the remaining details can identify someone.
- Use fictional examples when asking for a template, formula, email structure, or explanation. Add real facts later in the approved business system, not in the AI conversation.
- Check the output before using it. AI can produce confident but false statements, reveal bias, omit context, or invent citations. Verify important facts with authoritative sources and apply normal review and approval processes.
- Save only the final, reviewed work in approved storage. Follow retention, records, copyright, and attribution requirements. If an AI response contains material that looks private, proprietary, or unrelated to your prompt, stop and report it.
- If you entered prohibited information, do not try to hide or correct the incident by adding more details. Stop using that conversation, preserve the basic time and tool name, and contact support or your security contact promptly.
What not to do
Do not use a personal AI account to bypass workplace controls. Do not paste a whole document when a short fictional example will work. Do not ask an AI assistant to handle secrets or approve a high-impact decision. Do not assume that deleting a chat immediately removes every retained copy. Do not trust an AI-generated privacy statement, link, or instruction without checking the provider's official site and your organization's policy.
When to escalate
Contact support before proceeding when the approved tool or permitted data types are unclear. Escalate immediately if restricted information was submitted, an AI tool displays another person's information, you clicked a suspicious AI sign-in link, or you suspect account compromise. Use your organization's established security reporting channel for suspected phishing or exposure.
What to include in a ticket
Include the tool name, whether it was a work or personal account, the date and approximate time, the page or feature used, what type of information may have been exposed, and whether the prompt was submitted or only prepared. Include the exact error message and a screenshot only after redacting private or confidential content. Never include the exposed data itself, passwords, MFA codes, secrets, customer data, health data, financial data, source code, contracts, or confidential documents in the ticket.