Do not type passwords, one-time codes, full card or bank numbers, or government ID numbers into an AI agent chat. Every major agent has a safer route for those, such as a secure sign-in form, a vault or a takeover step where you enter the detail yourself. Health, legal and money details deserve a second thought, and other people's secrets are not yours to hand over. This list is based on the official docs and launch coverage as of November 2026.
Why chat is the wrong place for secrets
What you type into a chat is stored, can be added to memory, may be used for training depending on your settings, and can be reviewed by humans if flagged for safety. An agent also acts: it reads web pages and emails that could contain hidden instructions trying to get it to repeat what it knows. Our explainer on indirect prompt injection shows how that works.
- You type a secret
- Stored in chat history
- Saved to memory
- Exposed by injection, breach or review
The vendors' own credential systems exist to break that chain. Anthropic says "Claude Duet never sees your passwords, browser cookies, sign-in tokens, or other credentials" when you use its sign-in form. Typing a password into the chat bypasses that protection.
The never-share list
| Never share in chat | Why | Use instead |
|---|---|---|
| Passwords | Stored in history and memory | Vendor secure form, vault or password manager |
| 2FA and one-time codes | Gives full account access within minutes | Enter them yourself in a takeover step |
| Full card numbers and CVV | Payment fraud risk | Single-use agent cards or saved payment methods |
| Bank account and routing numbers | Enables transfers and fraud | Your bank's own app |
| Passport, driver's licence, SSN | Identity theft | Type into the official site yourself |
| Recovery phrases and seed words | Irreversible crypto theft | Nowhere; no legitimate agent needs them |
As of November 2026.
Several vendors enforce parts of this list. Duet's sensitive-info classifier never shares credentials, full card or bank numbers, or government IDs onward. Meta's Muse email connector filters out one-time passwords, password-reset links and magic links. OpenAI says saved passwords for dots go through a secure form and are not exposed to the model. Grok Bot hands passwords, 2FA, CAPTCHAs and payments back to you through a takeover flow, according to xAI's FAQ. These are safety nets, not reasons to paste secrets into a chat.
Think twice: health, legal and money details
These are not banned, but share only what a task needs.
- Health: Most consumer AI agents are not covered by HIPAA. Meta Muse trains by default unless you opt out, according to CNBC. Booking a physio appointment does not need your full diagnosis history. See AI agents and health data.
- Legal matters: Conversations with an AI agent are not protected by attorney-client privilege. Keep case details for your lawyer.
- Finances: Letting an agent track spending is different from giving it account numbers, and no mainstream agent needs your routing number to summarise your spending.
- Location and routine: Home address, children's schools and travel dates are useful to an agent and to a stalker. Share them only if you need the agent to act on them.
Work and account details
Work data is its own category. Client files, unreleased financials, source code and internal passwords may be covered by your employment contract or an NDA, and pasting them into a personal agent can breach both. Some vendors treat business plans differently: OpenAI says dots on Business, Enterprise and Edu plans are not used for training by default, while personal plans follow your ChatGPT setting. That does not make a personal account suitable for company secrets. Ask your employer which tools are approved before connecting a work inbox or drive.
Account recovery details deserve the same care as passwords. Security question answers, backup codes and recovery email addresses are effectively master keys. If an agent asks for one during a task, stop and take over the step yourself.
Other people's secrets
Your agent sees more than your own data. Connect your email or messages and it reads everything your contacts send you, including details they shared in confidence. Forwarding a friend's medical news or a colleague's salary to your agent means sharing it with a company they never chose. Our guide to bystander privacy covers group chats and contact data in more depth.
A practical test: if you would not forward the message to a stranger at a tech company, do not hand it to your agent.
How to share safely when you must
- Use the official credential flowSecure forms, vaults and takeover steps keep secrets away from the model.
- Share the minimumGive the agent a booking reference, not your full passport scan.
- Turn off trainingUse the parent account's model improvement setting.
- Prefer read-only connectorsRead access leaks less than send access; see our guide to read-only vs send access.
- Clean up afterwardsDelete the conversation and reset memory once the task is done.
Our guide to read-only vs send access explains how to scope connectors. For how agents keep passwords out of the model, see does your AI agent see your passwords.
What to do if you already overshared
- Change any password or PIN you typed, and turn on or reset 2FA.
- If you shared card details, ask your card issuer to replace the card.
- If you shared an SSN or ID number, consider a credit freeze with the three US credit bureaus and report identity theft at IdentityTheft.gov.
- Delete the conversation, then reset or edit memory using our memory deletion guide.
- Turn off training going forward; see opt out of AI agent training.
For the broader permission picture, see AI agent safety and permissions, and for codes arriving by text, AI text message scams.
Bottom line
Your AI agent should never need a password, code, full card number or ID number in chat. Use the vendor's secure flow or enter it yourself. For everything else, share the minimum for the task, keep other people's secrets out, and clean up afterwards. Pair this with our AI agent safety checklist for the full setup.



