uable.store
Able
← All articles How to Protect Sensitive Data from AI Models How-to

How to Protect Sensitive Data from AI Models

Last Updated: November 2, 2026

Quick Summary

To protect sensitive data from AI models, keep it out of prompts, anonymize what you must share, and verify each vendor's training and retention settings before submitting anything real. Cyberhaven reports 39.7% of AI interactions expose sensitive data, so a short pre-session checklist helps.

  • Cyberhaven reports that 39.7% of AI interactions expose sensitive data.
  • Perforce found 98% feel confident, yet 84% admit to compliance exceptions.
  • Never paste passwords, API keys, Social Security numbers, or full bank numbers into prompts.
  • Replace names, amounts, and dates with placeholders before sending any text.
  • Confirm training opt-out, retention window, deletion path, and subprocessors before using a tool.

Table of Contents

Sensitive data leaves your machine more often than most people realize. Cyberhaven reports that 39.7% of AI interactions expose sensitive data (Cyberhaven 2026 Report on sensitive data flowing into AI tools).

Able is a Windows desktop agent from Apex Empire Inc. It organizes files, runs terminal commands, and types vault logins only after you say yes. The model never sees the vault password.

Why Sensitive Data Leaks Through AI Tools Today

Most leaks happen by accident. Someone pastes a contract into a chatbot to get a summary, and the text is now on a third-party server.

Perforce's 2026 State of AI and Data Privacy Report found that 98% of respondents feel confident protecting sensitive data in AI workflows (Perforce 2026 State of AI and Data Privacy Report). Yet 84% also admit to compliance exceptions.

Common leak points include:

  • Pasting full documents into chat windows
  • Browser extensions that read page content
  • Copying logs or customer lists into prompts

Map the Data Flow From Prompt to Provider Storage

Before you protect anything, trace where your data goes. Follow one prompt from your keyboard to the end of its life.

Professional working at a desk, reviewing sensitive data on a laptop screen with a notepad nearby.
Professional working at a desk, reviewing sensitive data on a laptop screen with a notepad nearby.

A typical path looks like this:

  1. You type or paste text into an app.
  2. The app sends it over the internet to the model provider.
  3. The provider processes it and may log it.
  4. Logs may be kept for days, months, or longer.
  5. Some providers use inputs to improve future models unless you opt out.

Write this map for your own workflow. It takes ten minutes and shows you every place your data stops.

What Not to Put Into AI Prompts

Some data should never enter a prompt. Keep these out of any cloud tool:

  • Passwords, API keys, and login tokens
  • Social Security or national ID numbers
  • Full bank and card numbers

If a task involves this data, remove the sensitive parts first. The rest of the text usually works fine.

How to Anonymize Data for AI Before Sending It

Anonymizing means replacing real identifiers with placeholders before you paste anything.

Example:

Original: "Maria Lopez, 4412 Oak Street, owes $18,300 on account 77-204." Safe version: "Client A, [address], owes [amount] on account [ID]."

Good anonymization practice:

  • Swap names for roles such as "Client A" or "Vendor B"
  • Replace numbers with ranges or placeholders
  • Strip email addresses, phone numbers, and street addresses

Here is a quick template you can reuse:

"Summarize the following for [role]. Replace all names with [PERSON], all dollar amounts with [AMOUNT], and all dates with [DATE]. Text: [paste anonymized text]."

Checking Vendor Retention and Model Training Settings

Every AI vendor has its own retention and training policy, and most are written to be skimmed past. Verify four things before you submit anything real: whether your inputs train future models, how long prompts and outputs are retained, who the subprocessors are, and how deletion actually works. If you cannot answer all four in writing, treat the tool as public.

Map the data flow first

Before you read a single policy, draw the path one prompt takes. Most practitioners find five stops:

  1. Input capture, the chat box, browser extension, or desktop app that accepts the text.
  2. Transit, the TLS connection to the provider's API endpoint.
  3. Inference, the model that processes the prompt, plus any tool calls it triggers.
  4. Logging, request logs, abuse-monitoring stores, and evaluation sets kept by the provider.
  5. Storage, chat history, embeddings, and any files you uploaded.

Write this map for your own workflow on paper. Each stop is a place your data can be retained, and each needs its own answer. A prompt deleted from chat history can still live in an abuse-monitoring log for weeks; an uploaded file can persist in storage after the conversation is gone.

Verify training use, retention, and deletion

Ask the vendor, or read the policy, for these specifics, and write the answers down:

  • Training opt-out. Is there a documented setting, and does it apply to your plan tier? Free tiers frequently train on inputs by default; paid and enterprise tiers often do not.
  • Retention window. How long are prompts, outputs, and uploaded files kept? Look for a stated number of days, not "as long as necessary."
  • Deletion path. Is there a self-serve delete, an API delete endpoint, or only a support ticket? Test it on a throwaway prompt first.

Watch out: a vendor that says "we do not train on your data" but keeps prompts for 30 days for abuse monitoring has not given you a no-retention promise, those are two different settings, and both need checking.

Split the check by deployment type

The same checklist produces different answers depending on how you use AI:

Get Able →

  • Public chatbots (free or consumer tiers). Assume training-on-input is on unless the policy says otherwise. Never paste client records here. Use them for drafting and formatting only.
  • Enterprise SaaS seats. Retention and training terms are usually negotiated in the contract, not the public policy. Ask for the data processing addendum and confirm the training opt-out is contractual, not just a toggle.
  • AI agents on your PC. An agent that reads files and runs commands needs its own retention review. Able is a Windows desktop agent from Apex Empire Inc. that runs file, terminal, and vault jobs on this PC after you say yes; chat and research still go to cloud brains over the internet, so the provider-retention questions above still apply to anything you send to a cloud brain. Able is not Ollama or LM Studio, which run local models, Able is the Windows file agent that does the work on your PC. For the full split between what runs locally and what needs a brain, see Able's Windows desktop AI agent page.

Validate the settings with a test prompt

Policies are claims; tests are evidence. Run a short validation before you rely on any tool:

  • Send a prompt containing a unique marker string, then check whether it appears in chat history, exports, or logs you can access.
  • Delete the conversation and re-check the marker after the stated retention window.
  • Confirm the training opt-out is still set after a plan change or app update, settings sometimes reset.

For Able's own data-handling details, read the Trust page. For why the desktop-agent model fits this workflow, see why Able.

Before you submit anything real, check these four vendor settings and write the answers down.

Check What to confirm Why it matters
Training opt-out Documented setting and which plan tiers it covers Free tiers often train on inputs by default
Retention window Stated number of days for prompts, outputs and files "As long as necessary" is not a limit
Deletion path Self-serve delete, API endpoint, or support ticket Test it on a throwaway prompt first
Subprocessors Named third parties that touch your data A missing or stale list is a warning sign

Local AI Privacy: What Changes on Your Own PC

Running a model on your own machine keeps prompts from leaving the device. Local AI is not automatically private, though. Setup choices, plugins, and logs still matter. For a broader look, see Is Local AI Actually Private? 2026 Reality Check.

Able works differently from a pure local model. For a deeper guide on keeping AI files private on a Windows PC, read How to Keep AI Files Private on Your Windows PC. For the full picture of what Able does on the desktop, start with AI Desktop Assistant 2026: What Works on Windows PC.

Keep the split in mind:

  • Organize, find, and vault tasks run on your PC.
  • Chat and research send prompts to a cloud brain.
  • Only send what the cloud brain needs to see.

Run an AI Prompt Security Checklist Before Every Session

A checklist only works if it is short enough to use and specific enough to test. The version below is built around measurable controls, so you can prove it is working instead of assuming it is. Three controls, redaction, access scope, and revocation speed, catch most accidental exposures.

The pre-flight checklist

Run this before every session with any AI tool, including Able:

  • Remove passwords, API keys, and login tokens from the text
  • Replace names, IDs, and account numbers with placeholders
  • Confirm the vendor's training and retention setting for this session

Keep it to one screen. A checklist you skip is worse than no checklist, because it creates false confidence.

Turn the checklist into measurable controls

Each item maps to a control you can test and a number you can track:

  • Redaction control. Test with a sample document containing a fake name, a fake account number, and a fake email. After redaction, search the outgoing text for each marker. KPI: sensitive-data leakage rate, the share of test prompts where a marker survived redaction. Target zero.
  • Access-scope control. Give the agent a folder it should not read and confirm it is blocked. KPI: access violations per week. Target zero; any hit is a configuration bug to fix, not a one-off.
  • Revocation control. Time how long it takes to revoke a credential or remove a folder from the agent's allowed list. KPI: time to revoke. Minutes, not days.

Write the four numbers down once a month. A trend line is what turns a checklist into a security program.

What to do when something leaks

Even a good checklist fails sometimes. Have a short response playbook ready before you need it:

  1. Contain. Revoke the exposed credential or API key and remove the affected folder from the agent's allowed list. Do this first, before investigating.
  2. Preserve evidence. Export the conversation, the prompt, and any logs before deleting anything; you will need them for the notification decision.
  3. Assess scope. Identify which data classes were exposed, credentials, client records, financial data, and who is affected.
  4. Notify. Follow your own incident policy and any contractual notification deadlines. Do not wait for the vendor to notify you.
  5. Remediate. Fix the control that failed, re-run the test case, and update the checklist so the same path is covered next time.

Watch out: deleting the chat immediately after a leak feels like containment but destroys the evidence you need to decide whether notification is required, export first, then delete.

Where Able fits the checklist

Able handles several checklist items on this PC rather than relying on memory alone. Vault logins are typed on this PC after you say yes, and the model never sees the vault password. Terminal commands run after you say yes.

To test the setup on your own machine, start the 3-day trial with no card at uable.store/trial. For a closer look at scoping what an agent can touch, see How to Limit AI Agent Permissions on Your Windows PC. Able is sold as a one-time license; see Able pricing for current prices.

Set Up Able to Keep Vault Credentials Away From the Model

Credentials are the highest-risk data on any PC. Able keeps them out of the model's reach.

Setup steps:

  1. Install Able on Windows 10 or 11 from uable.store.
  2. Store logins in the vault during setup.
  3. Choose the folders Able can work in.
  4. Approve terminal commands before Able runs them.
  5. Start the free trial to test the setup before you buy.

The trial runs 3 days with no card required. Start at uable.store/trial and confirm the vault behavior on your own machine.

For a closer look at scoping what an agent can touch, see How to Limit AI Agent Permissions on Your Windows PC. For what is verified about Able's data handling, read the Trust page.

Protecting sensitive data from AI models comes down to controlling what leaves your machine, checking vendor settings, and keeping credentials out of prompts.

Frequently Asked Questions

How do I protect sensitive data when using AI?

Start by deciding what the model should never see: client identifiers, credentials, financial records, and unreleased internal documents. Replace names and account numbers with placeholders before you paste text, and turn off chat history or training options where the provider offers them. Check each vendor's retention policy, since some keep prompts for abuse monitoring even when training is disabled. Able handles vault logins separately: after you say yes, it types the credential on this PC, so the model never receives the password.

Can AI models see or store the information I enter?

It depends on the provider and the plan you use. Cloud AI services process your prompt on their servers, and many retain prompts for a set period for safety review or service improvement. Retention and training settings differ between consumer and business tiers, so read the data controls page for the exact product. Assume anything you type into a cloud chat may be stored unless the provider's documentation states otherwise. Able's chat and research features also use cloud brains, so the same caution applies to anything you send through them.

Is local AI more private than cloud AI?

Running a model on your own hardware keeps prompts off third-party servers, which removes one exposure path. It does not make everything private, because the model may still write outputs to logs, shared folders, or connected apps. Able is not a local model runner and is not fully offline: file organizing, search, and vault typing work on this PC without a brain, but chat needs internet and goes to cloud providers. For details on how Able divides the work, see https://uable.store/windows-desktop-ai-agent/.

What information should I never put into an AI prompt?

Avoid passwords, API keys, private keys, full card numbers, government ID numbers, medical records, unredacted contracts, and customer personal data. Also leave out source code that contains secrets and any document your employer classifies as confidential unless your approved AI tool is covered by a data processing agreement. When you need help with such material, describe the structure rather than the content. Able keeps vault credentials out of model context, so logins can be used on this PC without being pasted into chat.

GoogleAdd uable.store as a preferred source on Google →