Every artificial intelligence (AI) tool involves some risk. Each of us is in charge of our own risk appetite. This matrix collects the "gotchas" behind the Personal AI Security Checklist so you can see the issue, the everyday behavior that raises it, what could happen, and a lower-risk choice. Step IDs (O1, S1, P1, M1) point to the checklist.
Step 1: Know the level you're working at
These levels describe different kinds of access, not a universal ranking or measured risk score. Most people can do a lot at Level 1.
0
Setup: No AI
What it adds: Nothing new
What should be in place first: Nothing
1
Setup: Basic chat with your own, cleaned-up text. Web search, connections and agent actions off.
What it adds: Your text reaches the AI company and may be stored.
What should be in place first: Training off, memory reviewed, cleanup habits (S5, S6, P1 to P5)
2
Setup: Basic chat with outside content pasted in (emails, webpages, documents)
What it adds: Untrusted content that may hide instructions
What should be in place first: Treat it as untrusted, check answers, act yourself (P8 to P10)
3
Setup: Web search on
What it adds: Words from your conversation, and possibly memory, go to a search company
What should be in place first: Nothing sensitive in that chat (S8)
4
Setup: Connected to email, calendar, files or health records, read only, including AI built into your email app
What it adds: The AI reads things you haven't reviewed, including strangers' messages
What should be in place first: The three questions answered, scope limited, a plan to revoke access (S7, short-lived access checklist)
5
Setup: Connected, and able to act: send, share, book, buy, create events
What it adds: Actions you didn't intend, possibly triggered by outside content
What should be in place first: A tested off switch, supervision, approvals (S12, P11)
6
Setup: AI browser or agent working in your signed-in browser
What it adds: Potential access to signed-in accounts, subject to browser, site, and agent permissions
What should be in place first: A separate browser profile without banking, email or password manager (S11)
The risky combination: private data, untrusted content and a way out, all in one setup. Levels 4 to 6 can bring all three together.
Step 2: Know the gotchas
What you share and what's kept
Your storyline builds up
What's going on: Chats, memory and uploads add up to a profile of you, kept on the company's servers until deleted
Behavior that raises the risk: Sharing personal details across many chats, memory on
What could happen: A hacked account, data breach or change in terms exposes a detailed profile; some companies use chats for ads
Lower-risk choice: Share the minimum, review memory, close out each task
Steps: S6, P2, P14, M4
Training may be on by default
What's going on: Some services use your chats to improve models unless you opt out
Behavior that raises the risk: Never opening privacy settings
What could happen: Your content is kept longer and used in ways you didn't choose
Lower-risk choice: Turn off optional training
Steps: S5
Off isn't erased
What's going on: Training, memory and deletion are separate switches; deletion can take weeks; backups may keep copies
Behavior that raises the risk: Assuming one switch handles everything
What could happen: Data you thought was gone is still stored
Lower-risk choice: Check each setting; check retention rules
Steps: S5, S6, P14, M4
Signing out isn't private
What's going on: Without an account, you may not reach the opt-out settings
Behavior that raises the risk: Using AI signed out to "stay private"
What could happen: Chats may be used under default terms
Lower-risk choice: Use an account with your settings set
Steps: S5
Typed text may be sent before Send
What's going on: Some tools and writing helpers send text as you type
Behavior that raises the risk: Typing or pasting raw details, then deleting them
What could happen: Details reach the provider even if you never send
Lower-risk choice: Clean up text before it reaches the chat box
Steps: P3
Pseudonym labels aren't anonymous
What's going on: Details left behind can still point to a person; your label key reveals everyone
Behavior that raises the risk: Swapping names but leaving roles, dates and places; storing the key in AI
What could happen: People can still be identified; the key exposes them all
Lower-risk choice: Remove extra details; keep the key private
Steps: P4, P5
Files carry hidden details
What's going on: Comments, tracked changes, hidden sheets, author names and photo locations
Behavior that raises the risk: Uploading whole files
What could happen: Information you didn't see gets shared
Lower-risk choice: Check files; share only the needed part
Steps: P6
Other people's information
What's going on: Clients, patients, coworkers and family didn't agree to share
Behavior that raises the risk: Pasting their details to get help
What could happen: Broken trust, contracts or laws
Lower-risk choice: Confirm permission; share the minimum
Steps: P1, P2
Public share links
What's going on: Shared chats can be found; some have appeared in search results
Behavior that raises the risk: Making chats public, forgetting old links
What could happen: Personal details become public
Lower-risk choice: Keep sensitive chats private; search for yourself
Steps: P12, M5
Connections and outside content
Planted instructions
What's going on: Emails, invites and webpages can hide instructions meant for AI
Behavior that raises the risk: Asking AI to summarize your whole inbox or a webpage
What could happen: The AI exposes data or acts against you
Lower-risk choice: Treat outside content as untrusted; keep actions off
Steps: P8, P13
Ignored emails are ticking time bombs
What's going on: Spam and phishing you ignored still sit in your inbox
Behavior that raises the risk: Connecting AI to an inbox full of old junk
What could happen: A connected AI reads a planted instruction later
Lower-risk choice: Disconnect personal email; clear old junk if connected
Steps: S7
AI built into your inbox
What's going on: Email apps now include AI that can read any message, including strangers'
Behavior that raises the risk: Asking the built-in AI about mail from unknown senders, or about your whole inbox
What could happen: Hidden text pulls login codes or other details into the answer, or sends them out
Lower-risk choice: Use AI on email you chose to show it, not on email that chose you. Choose your level for built-in email, calendar and chat AI (Cautious, Balanced or Accepting), check even if you never turned it on, and keep AI from deleting or archiving mail
Steps: S7, S9, P13, M1
Long chats drift
What's going on: As a chat grows, the AI can lose earlier instructions and build on its own mistakes
Behavior that raises the risk: Running one chat for hours or across many topics
What could happen: Forgotten rules, mixed-up details and made-up answers
Lower-risk choice: One task, one chat; start fresh with a reviewed handoff summary
Steps: P15
Poisoned answers
What's going on: Scammers plant fake support numbers and links on websites so AI tools repeat them
Behavior that raises the risk: Calling a number or opening a link straight from an AI answer or summary
What could happen: You reach a scammer who asks for card details or codes
Lower-risk choice: Find contact details yourself in the official app, a bookmark or a statement
Steps: P9, P13
Calendar invitations
What's going on: Invitations from strangers can land on your calendar automatically
Behavior that raises the risk: Leaving auto-add on
What could happen: A connected AI reads a planted invitation
Lower-risk choice: Add invitations only after you accept them
Steps: S9
Data leaves without a click
What's going on: A loaded image, new event or shared file can carry data out
Behavior that raises the risk: Letting AI create or display content freely
What could happen: Data leaks with no warning
Lower-risk choice: Keep actions off; don't open links in AI answers
Steps: P9, P10
Web search is a door
What's going on: Searches send words from your chat, and possibly memory, to a search company
Behavior that raises the risk: Automatic web search on during private work
What could happen: Sensitive words reach another company under its terms
Lower-risk choice: Turn off automatic search for sensitive work
Steps: S8
Connecting copies your data
What's going on: Some connections copy data to another company
Behavior that raises the risk: Adding Gmail, Yahoo or AOL to another company's app
What could happen: More companies hold your data, under more terms
Lower-risk choice: Check who else gets a copy before connecting
Steps: O6
Middlemen keep copies
What's going on: Some connections run through a data service with its own terms
Behavior that raises the risk: Connecting health or financial accounts through a third party
What could happen: Copies outside your AI provider's control
Lower-risk choice: Prefer downloads; request deletion afterward
Steps: O6, short-lived access
Disconnecting isn't revoking
What's going on: Removing a connector in the AI app may leave permission at the source
Behavior that raises the risk: Disconnecting only in one place
What could happen: Access continues
Lower-risk choice: Revoke at the source account too
Steps: S7, M6
Read-only still copies
What's going on: The tool gets copies of whatever it can read
Behavior that raises the risk: Assuming read-only means safe
What could happen: Private data is exposed even without changes
Lower-risk choice: Limit what it can read
Steps: O5
Health information hides everywhere
What's going on: Inboxes hold appointment, pharmacy and insurance messages
Behavior that raises the risk: Connecting email without thinking about health details
What could happen: Health conditions inferred and stored
Lower-risk choice: Disconnect personal email; use the Cautious path
Steps: S7, health note
No BAA on personal plans
What's going on: HIPAA coverage depends on the recipient's role and whether the app acts for a covered entity or business associate
Behavior that raises the risk: Using a personal AI account for patient or client data
What could happen: Legal and contract violations; harm to others
Lower-risk choice: Use an approved business account, configured correctly
Steps: P1
Accounts, actions and tools
Your AI account is a master key
What's going on: A connected AI account can reach chats, files and connected services
Behavior that raises the risk: Weak passwords, no multifactor authentication
What could happen: One takeover exposes everything connected
Lower-risk choice: Passkey or MFA; protect recovery
Steps: S1 to S3
Agents act
What's going on: AI that can send, buy or share can do it wrong, or be steered
Behavior that raises the risk: Letting an agent run unsupervised
What could happen: Messages sent, files shared, money spent
Lower-risk choice: Act yourself; stay nearby
Steps: P10, P11
Hard to stop
What's going on: Stopping a task, a schedule and access are different controls
Behavior that raises the risk: Not knowing where the off switch is
What could happen: Actions keep running; stopping doesn't undo
Lower-risk choice: Find and test the off switch first
Steps: S12
AI browsers use your sign-ins
What's going on: An AI browser acts inside accounts you're signed into
Behavior that raises the risk: Using it in your everyday browser profile
What could happen: Account takeover, data theft
Lower-risk choice: Separate profile without banking, email or password manager
Steps: S11
Installs that stay
What's going on: Plugins, skills, extensions and add-ins keep their access after the task; a document add-in may read document contents, depending on its permissions
Behavior that raises the risk: Clicking Install without knowing what it is, where it goes or what it can read
What could happen: Data theft, hidden copies or unwanted actions
Lower-risk choice: Know what, who, why, where, what it reads and how to remove it, or copy plain text into a basic chat instead
Steps: S10, P6
Confidently wrong answers
What's going on: AI can be wrong and sound sure
Behavior that raises the risk: Relying on totals, dates or advice without checking
What could happen: Bad decisions, especially on health and money
Lower-risk choice: Check against originals and sources
Steps: P9
Updates change things
What's going on: New features and defaults can arrive with updates
Behavior that raises the risk: Setting things once and never looking again
What could happen: Search, memory or connections back on
Lower-risk choice: Check monthly and after updates
Steps: M1
Forgotten accounts
What's going on: Old AI accounts still hold your data
Behavior that raises the risk: Signing up and moving on
What could happen: Exposure in a breach you won't notice
Lower-risk choice: Find and close unused accounts
Steps: O2, O3, S13
Step 3: Decide your risk appetite
For any task, ask three questions:
- How sensitive is the information? Low (a recipe idea), medium (a household budget), high (health, finances, children's routines, other people's data).
- What happens if the answer is wrong? Low (a meal idea), medium (a missed renewal), high (a missed pickup, a medication question, a legal deadline).
- What level from Step 1 does the task need?
Then match them:
Low and low
Suggested level: Up to Level 3 is reasonable, with care
Medium on either
Suggested level: Level 1 or 2, with cleanup and checking
High on either
Suggested level: Level 1 only, with the minimum information, or do it without AI
High information and you still want a connection
Suggested level: Short-lived access only, using the Short-Lived Access Checklist
These are starting points, not rules. Your comfort level, your obligations to other people and your work rules may call for more caution.
Changing a setting or disconnecting doesn't necessarily erase data already received. A checked box records a decision, not a guarantee of safety. General information, not legal advice. The author develops commercial security tools in this area; this matrix is tool-neutral and recommends no product. CC BY 4.0. Provided as is, without warranty.