You asked ChatGPT to rewrite a client email. You dropped a contract into Copilot so it could summarise the clauses. You pasted a stack trace that still contained an API key. None of that feels like handing over company secrets. To the AI provider, it is still data leaving your organisation.
IBM's 2025 Cost of a Data Breach report found that one in five organisations studied had a breach involving shadow AI: unsanctioned AI tools used without IT oversight. Only 37% had policies to manage AI or detect that usage. Samsung's 2023 ChatGPT incidents showed how quickly this happens in real work. Engineers pasted proprietary source code and meeting notes into a public chatbot to save time.
This guide is for employees. You do not need to become a security engineer. You need one habit: treat every prompt as if a stranger could read it later.
Treat every prompt like a public channel
Consumer AI tools send your prompt to servers you do not control. The vendor may store it, use it to improve models, or keep it for abuse review. Enterprise tools with a contract and "do not train" settings are safer. They are still not a password vault. If you would not put it in an unencrypted email to a vendor, do not paste it into a prompt.
Related reading: Using AI: the impact on (your) data privacy.
What counts as a secret at work
Anything that would hurt customers, colleagues, or the company if it left the building:
- Passwords, passphrases, recovery codes, MFA codes, session cookies
- API keys, access tokens, certificates, connection strings, SSH keys
- Customer names with account numbers, invoices, contracts, health or HR files
- Source code, architecture diagrams, unpublished product plans
- Internal meeting notes, incident reports, salary data
- Anything labelled confidential, restricted, or personal data under GDPR
If you are unsure, treat it as confidential. Data classification exists for this reason.
Practical tips for employees
1. Never paste credentials
Do not ask an AI to "check if this password is strong," "debug this login," or "store this key for later." Paste a redacted example instead: API_KEY=REDACTED. Keep real secrets in a company password manager or digital vault, not in a chat window.
2. Strip names and identifiers before you ask for help
Replace real values with placeholders: CUSTOMER_A, IBAN_REDACTED, server-1.internal. You can still get a useful rewrite, summary, or code suggestion without handing over personal data.
3. Do not upload files you would not email to a stranger
PDFs, slide decks, spreadsheets, recordings, and screenshots are prompts too. A "summarise this contract" upload can include pricing, personal data, and legal strategy in one click. Share files through approved channels instead. See secure document and file sharing best practices.
4. Use only company-approved AI
A personal ChatGPT account on your phone is still a company leak if the content is work data. "Shadow AI" is the name for this pattern: helpful tools that IT never approved. Follow your AI policy. If there is no policy yet, ask before you paste.
5. Turn off training and history when the tool allows it
Free and consumer accounts often use chats to improve models unless you opt out. Team and enterprise plans usually do not train on your prompts by default, and they add admin controls. Settings help. They do not make a password safe to paste.
6. Watch meeting bots and AI notetakers
Transcription tools can capture passwords read aloud, incident details, and customer names. Check who owns the recording, where it is stored, and whether guests can join without approval. Do not invite an unknown notetaker into a confidential call.
7. Do not ask AI to remember or store secrets
Memory features, custom instructions, and "save this for next time" are convenient. They also create a second copy of the secret outside your vault. If a colleague needs a credential, share it through the vault or a secure link, not through a chatbot.
8. Be careful with browser extensions and "free" AI helpers
Unvetted extensions can read the page you are on, including password fields and internal tickets. Install only tools your company allows. Treat a random "AI sidebar" like any other software that can see your screen.
9. Keep source code and configs clean
Samsung's leaks started with a reasonable request: fix this code, summarise these notes. Before you paste a snippet, search it for keys, tokens, hostnames, and comments that name customers. Prefer dummy data in examples.
10. Report quickly if you already pasted something
Speed matters more than embarrassment. Tell IT or security what you pasted, which tool you used, and whether the chat is still open. They can rotate keys, revoke tokens, and contain the issue. Hiding it turns a mistake into a data breach that lasts longer. If you need the organisational playbook, read responding to a data breach.
A 30-second check before you hit send
| Question | If the answer is no, stop |
|---|---|
| Would I put this in a public Slack channel? | Remove it or redact it |
| Does this include a password, key, or token? | Use the vault instead |
| Does this identify a customer or colleague? | Replace with a placeholder |
| Is this the approved AI tool? | Switch or ask IT |
| Could this file contain hidden sheets or comments? | Check before you upload |
If you already shared a secret
- Stop using that chat for anything else related to the secret.
- Copy the conversation ID or URL if the tool shows one, then close it.
- Report to IT or security with the time, tool, and what was pasted.
- Rotate the credential or key. Do not reuse the old value.
- Follow your incident process. Do not try to "delete it from the internet" yourself.
You cannot always pull a prompt back from a vendor. You can cut off the value of what leaked by rotating access.
Make the safe path the easy path
Employees paste secrets into AI because the alternative is slow: hunting through chat history, asking a colleague, or waiting on IT. A shared vault fixes that. Store logins, API keys, and files where they belong. Share them with the right people under access control, not in a prompt.
Hypervault is a European password manager and digital vault with zero-knowledge encryption. Teams keep secrets out of chat tools and AI prompts, then share them on purpose. Compare plans or start a trial.
