Keep passwords, API keys, recovery codes, private keys, and wallet seed phrases out of DeepSeek prompts. Leave out sensitive personal records and confidential work material that you are not authorized to send to the service you are using. The same rule applies when the information is inside a screenshot, attachment, code sample, or earlier message in a conversation.
You can still get useful help with an email, document, or coding problem. Usually, the model needs the task, relevant facts, and constraints. It does not need the customer’s identity, the real access token, or the whole internal file. This guide shows how to keep the useful context, remove unnecessary exposure, and respond if you have already shared something sensitive.
Already pasted a secret? Go straight to what to do after sharing sensitive information. Deleting a conversation does not invalidate a working password or API key.
On this page: What to leave out · Which service you use · How to prepare a prompt · Practical examples · Files and screenshots · Shared chats and agents · If you already shared it · FAQ
What not to paste into DeepSeek
Start by separating information that grants access from information that needs permission to disclose. A real password has no useful role in a request for login troubleshooting. A confidential contract might be useful to analyze, but its usefulness does not establish permission to upload it.
| Information | What to leave out | What to use instead |
|---|---|---|
| Passwords and access secrets | Passwords, one-time codes, recovery codes, API tokens, session cookies, private keys, database credentials, and wallet seed phrases. | A placeholder and a description of the problem. Enter credentials only in the appropriate trusted authentication or configuration interface. |
| Identity and private contact details | Passport or national ID numbers, full addresses, personal phone numbers, account identifiers, signatures, and identity-document images. | Role labels such as CUSTOMER_A, with only the context needed for the task. |
| Health, financial, and highly personal records | Medical reports, payment-card details, bank or tax documents, biometric information, intimate material, and identifiable information about children. | A general question, fictional example, or a service specifically approved for the information and purpose. |
| Customer, employee, and student records | Support exports, payroll, CV collections, performance reviews, disciplinary cases, grades, and attendance or case histories. | A synthetic sample or a general description of the workflow. Use real records only through an authorized process suitable for that data. |
| Confidential business and legal material | Unreleased plans, internal pricing, board papers, trade secrets, client deliverables, private contracts, and communications with legal advisers. | A non-confidential question or approved excerpt. Ask the responsible owner when the substance itself is restricted. |
| Private code, logs, and security findings | Production configuration, secrets in error logs, customer payloads, proprietary logic, private infrastructure details, and undisclosed vulnerabilities. | A minimal example using invented data, preserving only the behavior needed to explain the issue. |
For the services covered by its Privacy Policy, DeepSeek explicitly asks users not to provide sensitive personal data about themselves or others. A person’s permission to discuss their information does not remove that service restriction.
Confidentiality also extends beyond personal information. Removing names from an unreleased acquisition plan still leaves an unreleased acquisition plan. If the business decision, algorithm, or contract terms are the secret, a few placeholders will not make the document suitable for an ordinary chatbot.
This does not make all work-related writing off-limits. Editing an approved public announcement is different from uploading the internal discussion behind it. Ask whether the actual content you intend to send is permitted, necessary, and appropriate for the chosen service.
Check which DeepSeek service will receive the information
The DeepSeek name can refer to the official chat service, its developer API, or a model used inside another application. Identify the website or app operator and the selected provider before relying on a privacy setting.
Official DeepSeek chat and app
DeepSeek’s Privacy Policy, dated February 10, 2026, covers services that link to it. It describes collecting inputs, uploaded files, photos, and chat history, and processing and storing personal data in China. It excludes end-user processing in downstream applications built through the Open Platform.
Section 4.3 of the Terms of Use describes using inputs and outputs to improve the service under stated safeguards, with an opt-out named Improve the model for everyone. Treat this as a control over that use. It is not a promise that the service no longer receives the prompt or retains no records.
For ordinary questions, keep the input non-sensitive regardless of that choice. Our DeepSeek privacy policy guide explains the collection, training, and deletion provisions in more detail.
API integrations and third-party apps
An API key or a paid plan does not, by itself, establish that a workflow is approved for client information. The Open Platform terms require developers to have appropriate input permissions and, for downstream services, disclose personal-data processing, establish a lawful basis, and address users’ rights. They distinguish a developer’s own personal information from downstream end-user processing.
If you are using an employer’s integration, ask which data categories are approved and whether the approval covers this particular provider and use. If you operate the integration, check what your application, model provider, logging tools, and connected services receive. An otherwise appropriate provider cannot compensate for an application that sends the whole customer record unnecessarily.
Local models and connected tools
A model running on your own computer can avoid sending inference requests to an external model provider. Check the application around it, including remote fallback, saved histories, synchronization, and connected tools. “Open source” describes availability and licensing; it does not establish where a particular app sends data.
For example, the DeepSeek Harness Privacy Policy distinguishes official model services from custom model services. With a custom model, inputs go directly to the selected provider; third-party tools have their own data handling. A local interface can therefore still be part of a cloud workflow.
For work material, use the deployment your organization has approved for that class of information. Our data residency guide covers how to check the route and storage locations. Here, the immediate decision is what the task actually needs you to send.
How to prepare a useful prompt without oversharing
Do the cleanup before uploading or pasting. Asking an online chatbot to “anonymize this confidential document” still sends it the original document. Start with a separate copy or write a short description yourself.
- Define the task. Are you asking for a friendlier reply, an explanation of a clause, or help with an error? Choose the smallest passage or example that supports that task.
- Remove identifiers and secrets. Replace names, addresses, account details, tokens, and private URLs consistently. CUSTOMER_A should mean the same person throughout; another person should have a different label.
- Review the remaining context. A rare job title, precise location, date, or distinctive incident may still identify someone. Generalize unnecessary details or switch to a fictional scenario.
- Preserve what affects the answer. Keep relevant relationships, units, conditions, column types, error codes, and sequences. If values are sensitive, use invented values that preserve the issue and label them as invented.
- Read the final version on its own. Check that it both explains the problem and omits the restricted information. Keep any mapping back to real identities outside the prompt.
Replacing names is not a guarantee of anonymity. The UK’s ICO guidance on pseudonymisation distinguishes replacing identifiers from making people no longer identifiable. If other details or a separate lookup can link a record to someone, treat the result cautiously rather than declaring it anonymous.
Sometimes the safe version no longer supports the original task. A generic account cannot settle a confidential contract dispute, and a fictional dataset cannot establish facts about real customers. In that case, change the request to a general explanation or use an authorized environment for the real material. Do not quietly put the sensitive details back to improve the answer.
Four examples that keep the question useful
The examples below are fictional prompt patterns, not tests or real customer records. They show which details carry the task and which can stay out of the conversation.
1. Rewrite a customer email without the customer record
Instead of pasting an entire billing thread with a name, address, invoice number, payment details, and previous complaints, describe the action and the tone you need:
Write a courteous reply to CUSTOMER_A about a duplicate charge. The refund has been approved. Explain that processing is expected to take 5–7 business days. Ask the customer to contact support if it has not arrived after that period. Do not request card details. Use [SUPPORT_CHANNEL] as a placeholder.
The processing time here is an illustrative assumption. Replace it with your verified, non-confidential policy. DeepSeek has enough context to draft the reply; the customer’s payment card and invoice history would add exposure without improving the writing task. Add the real greeting and approved contact details in your own email system afterward.
2. Ask about a contract concept without sending the contract
For a general explanation, rewrite the relevant concept in your own words. Keep out party names, signatures, confidential prices, settlement details, and privileged legal advice:
Explain this fictional service-contract arrangement in plain English: Either party may terminate with 30 days' notice. Outstanding fees remain payable, and confidentiality duties continue afterward. Explain the practical questions someone should clarify before relying on this wording.
This can help you understand notice, remaining obligations, and ambiguous wording. It does not establish the meaning or enforceability of a real contract. If you need advice on an actual agreement, involve the responsible legal adviser and use their approved process for sharing it.
3. Share a small coding example without credentials or production logs
Keep the language version, relevant function, input shape, and exact error type. Remove authentication headers, cookies, environment files, connection strings, private hostnames, customer payloads, and unrelated stack-trace paths.
Help me understand this Python error using invented data.
Input: [{"amount": "12.50"}, {"amount": "7.25"}]
Code: total = sum(row["amount"] for row in rows)
Error: TypeError when adding an integer and a string.
Expected result: a numeric total of 19.75.
Explain an appropriate way to handle these decimal values.
The example retains the important detail: the amounts are strings. Replacing every value with “[REDACTED]” would hide the cause. For your own issue, check locally that the reduced example still represents the problem before sharing it; do not include proprietary logic merely to make the snippet look complete.
Using an API key in a trusted client’s credential setting is separate from pasting it into the model’s conversation. DeepSeek’s Open Platform terms, section 2.2, specifically warn against disclosing keys or exposing them in browser or other client-side code. A troubleshooting prompt can use [API_KEY] instead of the real value.
4. Explain a spreadsheet problem with synthetic rows
If you need a formula, recreate the relevant columns in a clean sheet with invented rows. For a monthly sales formula, that might mean a date, a category, and an amount. Customer names, addresses, payment references, and the rest of the workbook are usually unnecessary.
Keep the features that make the problem difficult: a blank cell, a duplicated row, a date stored as text, or a category with no match. State the expected result for the invented example. If the task is to analyze real performance, invented data cannot supply the real conclusion; use the sample to design the method, then apply and check it in your approved environment.
Check files and screenshots beyond what is visible
A file can contain more than the page you are looking at. Copying a short, reviewed passage into a new document is often easier to assess than cleaning an entire report. When the file itself is necessary, inspect a separate copy and open the final version before attaching it.
Documents, spreadsheets, and presentations
Check comments, tracked changes, document properties, hidden text, hidden sheets, and presentation notes. In supported desktop Office apps, Microsoft’s Document Inspector guidance uses File → Info → Check for Issues → Inspect Document. Run it on a copy and review what it finds.
Some findings need manual attention. Removing hidden spreadsheet data can also change formulas or results. Recheck the cleaned copy, including the values needed for your question, rather than treating an inspection result as permission to share the whole workbook.
PDFs and scanned documents
Use an actual redaction function for content you need to remove. Drawing a colored rectangle over text is not the same operation. In Acrobat Pro, Adobe’s redaction instructions require selecting the content, applying the redaction, and saving the result.
Visible text and hidden information need separate attention. Adobe’s PDF sanitization guidance covers metadata, comments, hidden layers, and other embedded information. Reopen the saved copy and inspect the relevant pages. A visual check alone cannot prove that every hidden element has been removed.
Screenshots, images, and recordings
Capture only the area needed to explain the problem. Check account menus, browser tabs, addresses, notifications, QR codes, names in sidebars, and tokens visible in developer tools. Review the exported image itself so you know which version you are attaching.
For audio or meeting transcripts, review participant names and the substance of the discussion as well as any attached notes. A transcript of an internal strategy meeting can remain confidential after names are removed. Use a fictional summary of the task if permission to share the actual discussion is unclear.
Finally, check filenames and attachment selections. A sanitized document named after a client or case can reintroduce information you removed from its contents. Avoid uploading both the cleaned copy and the original.
Look beyond the latest prompt: shared chats and agents
Review the whole conversation before sharing it. The last answer may be harmless while an earlier message contains a private example or an attachment you forgot about.
DeepSeek’s Terms of Use, section 1.6, say anyone with a shared link can view the linked dialogue. They also warn that public republication can expose it to crawlers despite anti-crawling measures. A link intended for one colleague can be forwarded, so create a clean summary for sharing when the original thread contains material that should stay private.
Treat instructions inside other people’s content as untrusted
Summarizing a webpage or document is a normal use case. The additional risk is that the content may contain instructions intended to redirect the assistant, including instructions hidden from ordinary view. OWASP’s prompt injection guidance explains that external text, files, and images can influence behavior, with greater consequences when the assistant can use connected systems.
Keep untrusted material separate from sensitive conversations and give connected tools only the access needed for the task. For a document summary, avoid granting unrelated access to email, private folders, or publishing functions. Check proposed actions before allowing a tool to send, change, or delete anything.
You can label pasted content as material to summarize and tell the assistant not to follow instructions inside it. That is a useful direction, but it is not a security boundary. Restricting actual permissions and reviewing actions matter more than expecting one sentence to block every malicious instruction.
What to do if you already pasted sensitive information
Stop adding details to the conversation and identify what was exposed, which service received it, and whether a shared link or connected tool widened access. The response depends on whether the information can grant access, identify someone, or reveal confidential work.
If it was a password, API key, or other access secret
- Invalidate the exposed credential through the issuing service. Revoke or rotate an API token; change an exposed password using the service’s official security controls. Invalidate related sessions or recovery codes when applicable. For a workplace credential, contact its owner or security team promptly.
- Update legitimate uses. Replace the old credential in approved secret storage and dependent applications. Verify that the replacement works and the exposed value no longer grants access.
- Review activity. Check available sign-in, access, usage, and billing records for unexpected activity. Escalate suspicious use to the service or your security team.
- Reduce the remaining exposure. Follow the service’s available conversation, file, and sharing controls. Keep a record of the incident without copying the live secret into more chats or tickets.
Revocation comes first because removing a copy does not disable a credential. GitHub gives the same priority in its guidance on exposed repository secrets. The storage location differs here, but the practical point carries over: make the exposed value unusable.
A wallet seed phrase needs the wallet provider’s official recovery guidance; changing an app password does not replace the seed. Do not paste it into another chatbot or send it to someone offering recovery help.
If it was personal data or confidential work
Record the service, account or workspace, approximate time, type of information, and any sharing or tool actions involved. For work, client, patient, student, or employee information, notify the responsible privacy or security contact promptly. Follow their process for preserving evidence and containing the exposure.
Use the relevant service’s removal controls or request route. DeepSeek’s Privacy Policy describes chat-history deletion through settings and rights requests through privacy@deepseek.com, subject to applicable rules and exceptions. It describes variable retention rather than a universal deletion deadline. For a third-party app, contact its operator about its copies and downstream handling.
Describe the incident in a support request without unnecessarily attaching the sensitive material again. Deletion or withdrawal of a shared link cannot retrieve copies another recipient already made. If files were uploaded separately, check their controls too instead of assuming removal of a conversation covers every stored item.
Do not decide on your own that no further action is needed because the assistant says it has “forgotten” the information. Equally, sharing does not automatically establish that an attacker accessed it. The responsible privacy or security team should assess the facts, contractual duties, and any notification requirements.
A prompt structure you can reuse
For routine, non-sensitive work, use a structure that forces you to choose the context before submitting it. Fill this out with reviewed information rather than attaching the original document by default:
Task: [What I need help with] Audience or purpose: [Who the answer is for] Relevant context: [Only the facts needed for this task] Example: [A reviewed excerpt or clearly labeled fictional sample] Constraints: [Tone, format, rules, or technical conditions] Expected result: [What a useful answer should let me do] Use the placeholders as written. If information is missing, ask rather than inventing it.
A general outline, public documentation you are entitled to use, a fictional scenario, or a small non-sensitive example is often enough. Check the answer before use, especially when placeholders or changed details affect the conclusion. Add real identities or restricted values back only within the appropriate environment.
Frequently asked questions
Can I upload my CV or a job application?
For wording feedback, a copy with relevant experience and skills may be enough. Remove private contact details, full address, identity numbers, references’ contact information, and confidential information about employers. A distinctive career history may still identify you, so decide whether sharing the remaining content is appropriate. Do not upload other applicants’ CVs simply because you can access them at work.
Is private source code always off-limits?
No single rule covers every authorized development environment. The decision depends on ownership, confidentiality obligations, organizational approval, the selected service, and what the code reveals. For an ordinary chat, prefer a minimal example you are permitted to disclose. Keep real credentials out of the prompt even when the code itself is approved.
Can I ask about a medical or financial term without uploading records?
A general educational question can be asked without a report, bank statement, or personal case history. Ask what a term means or use a fictional example. Keep sensitive records out of the consumer service, and use an appropriate professional for decisions about your actual circumstances.
Does turning off model improvement make confidential prompts safe?
It addresses a specific use of data, not all collection, processing, storage, sharing, or access. You still need permission to submit the information and a service suitable for it. Do not treat the setting as approval for client records, workplace secrets, or sensitive personal data.
Can I ask DeepSeek to redact a document for me?
Only if sending the original to that service is already appropriate and authorized. Otherwise, the first upload creates the exposure you wanted to avoid. Prepare a clean copy beforehand, or use an approved local process that does not send the original elsewhere.
Does incognito mode or a VPN prevent DeepSeek from receiving my prompt?
No. The online service still receives the content you submit so it can respond. Google explains that Incognito limits what Chrome saves on your device; it does not stop websites collecting information. A VPN likewise does not remove the text you send to the destination service.
Can I use a local DeepSeek model for confidential documents?
A genuinely local workflow can keep model requests on your infrastructure, but you still need permission and appropriate controls over the device, application, files, histories, and tools. Check for remote providers or synchronization before assuming the whole workflow is local.
Is deleting the chat enough after a mistake?
It can be part of cleanup, but the next step depends on the information. Revoke exposed credentials, notify the responsible contact for work or other people’s records, and check separate files or shared links. Do not assume deletion reverses processing or removes copies outside the service.
Before you send the next prompt
Choose one task and prepare the smallest useful version of its context. Keep the details that explain the problem, replace unnecessary identifiers, and leave secrets out entirely. If the answer depends on confidential material you cannot appropriately share, use an approved workflow or ask a more general question. That gives you a practical way to keep using DeepSeek without turning every useful request into a disclosure of the original record.