Privacy and Security When Using AI

Protect people, accounts, confidential information, and organizations before data enters an AI system.

Personal + workplace15 minute readReviewed July 24, 2026

Know Where the Data Can Go

An AI interaction may involve a website or app, model provider, cloud host, plugins or connectors, logging and safety systems, organization administrators, and external tools. “Not used to train the model” does not necessarily mean “never stored,” “never reviewed,” or “not processed by another service.” Terms differ by product, plan, region, and setting and can change.

Prompts and uploads

Text, images, audio, documents, metadata, and hidden document content can contain identifiers or confidential material.

Memory and history

Saved chats and personalization can make work convenient but extend retention and expose information to anyone who gains account access.

Connected tools

Email, drives, calendars, code repositories, and agents expand what an AI system can read or change. Scope access narrowly.

Generated output

Outputs may accidentally reveal supplied data, include secrets copied from source files, or be shared more broadly than intended.

Data minimization is the strongest default: if the model does not need a piece of information to perform the task, do not provide it.

Before You Share: REDACT

R — Remove names, faces, addresses, signatures, precise locations, and unnecessary metadata.
E — Exclude secrets: passwords, API keys, recovery codes, private keys, session cookies, and security answers. Rotate a secret immediately if exposed.
D — De-identify records with placeholders, but remember that combinations of details can re-identify a person.
A — Ask who owns it. Confirm permission for client, employer, student, patient, copyrighted, or partner material.
C — Check settings and contract: retention, deletion, model training, subprocessors, region, admin controls, and export options.
T — Test with synthetic data before connecting real systems or sensitive documents.

Never put in a general-purpose public AI chat	Use only with explicit approval and controls	Usually lower risk
Passwords, keys, financial account credentials, identity documents, intimate images, active security details	Health, legal, HR, student, customer, proprietary code, contracts, unpublished financials, personal records	Public information, fictional examples, synthetic records, your own non-sensitive drafts

“Anonymous” is not automatic. Remove direct identifiers and distinctive combinations such as exact age, rare diagnosis, workplace, date, and village.

Guidance by Setting

Families and children

Use age-appropriate supervision. Do not create public digital histories for children by uploading private conversations, school records, faces, voices, or location routines. Teach children that AI is not a confidential friend.

Schools

Use approved accounts and published policies. Protect student records and accommodations. Provide non-AI alternatives and do not infer ability, misconduct, or emotion from opaque automated scores.

Nonprofits and communities

Donor, beneficiary, immigration, health, and safeguarding information can create serious harm if exposed. Use consent, role-based access, minimal retention, and a documented review process.

Small businesses

Maintain an approved-tool list and data-classification rules. Require MFA, separate accounts, offboarding, vendor review, and human approval before agents send messages, make purchases, or alter records.

Consent must be meaningful. Someone sharing information with you did not automatically agree to have it uploaded to an AI provider.

Security Risks That Grow With AI

Risk	What it looks like	Defense
Phishing and impersonation	Polished personalized messages, voice clones, fake support	Verify independently; use MFA or passkeys; never share one-time codes
Prompt injection	A document or webpage tells an AI agent to ignore instructions, reveal data, or take an action	Treat external content as untrusted; isolate tools; allowlist actions; require human approval
Excessive agency	An agent has broad email, file, payment, shell, or admin privileges	Least privilege, read-only defaults, spending limits, sandboxes, logs, and confirmation gates
Supply-chain risk	Untrusted model files, plugins, extensions, packages, or remote code	Use reputable sources, verify hashes/signatures, scan, pin versions, isolate execution
Secret leakage	Credentials appear in prompts, logs, repositories, screenshots, or output	Secret managers, automated scanning, redaction, restricted logging, immediate rotation

For developers, do not place authorization decisions in a prompt. Enforce permissions in deterministic application code, validate tool arguments, encode output safely, and assume retrieved content may be hostile.

If You Shared Something You Should Not Have

Contain

Stop sharing, revoke links and connectors, disable affected automation, and preserve necessary incident details.

Rotate

Change exposed passwords and revoke API keys, sessions, recovery codes, tokens, or certificates. Deleting a chat is not enough for a leaked secret.

Report

Notify your security or privacy contact and the provider. Follow applicable organizational and legal reporting requirements promptly.

Learn

Determine what was exposed, who could access it, retention settings, and the control needed to prevent recurrence—without blaming the reporter.

Five-minute privacy practice

Open one AI account you use. Review chat history, memory, training/data-control, connected apps, shared links, active sessions, and multi-factor authentication. Remove anything unnecessary and record which tasks are approved for that account.