Alex has three tasks to finish: improve a CV, reply to a customer and summarise a work document. An AI assistant could help with all three. But just before uploading the CV, Alex notices a home address, phone number and a referee’s contact details.
Does the assistant need any of that information to improve the writing?
That pause is where AI security for beginners starts: understanding what you share, what a tool can access and what needs checking before you act.
Alex is a hypothetical beginner. The connected scenarios in this article are fictional examples, not accounts of personal testing or real customer experiences.
This guide explains how to use AI tools safely through everyday decisions, without technical expertise.
Table of Contents
- What Is AI Security?
- Everyday Risks and Practical Precautions
- Protect Sensitive Information: Alex’s CV
- Check the Tool, Privacy Settings and Account Security
- Reply to Customers Without Oversharing
- Summarise Work Documents and Limit Permissions
- Check AI Answers Before Trusting Them
- Recognise AI Scams
- Respond to Accidental Sharing or a Scam
- Your 10-Minute AI Safety Review
- Frequently Asked Questions
What Is AI Security?
AI security involves protecting AI systems, their users and their information from misuse, unauthorised access and harmful manipulation. For everyday users, it includes choosing legitimate tools, limiting data exposure and controlling what an assistant can do.
Generative AI creates content such as text, images or audio. A prompt is the request or instruction you give it. Some tools only generate responses; others connect to files, email or services and perform actions.
Privacy and security are different
Privacy concerns how your information is collected, used, retained and shared.
Security concerns protecting information and systems against unauthorised access, alteration or disruption.
A service might protect its accounts well while still using conversations in ways you do not expect. You therefore need to check both its security controls and its data practices.
Everyday Risks and Practical Precautions
| Risk | Everyday example | Practical precaution |
| Unnecessary data sharing | Uploading a complete CV for a paragraph edit | Share only the relevant, edited section |
| Hidden information in uploads | A screenshot includes a customer’s address | Inspect the whole image before sending |
| Incorrect answers | A reply invents a delivery date | Verify against the original record |
| Fake apps | An imitation assistant requests your login | Reach downloads through the provider’s official website |
| Excessive permissions | A writing tool requests access to all email | Decline access that the task does not require |
| Malicious instructions | A document tries to redirect the assistant | Restrict actions and review unexpected behaviour |
| Impersonation scams | An urgent payment request sounds familiar | Verify through a separate, trusted contact method |
Protect Sensitive Information: Alex’s CV
Alex’s first task is improving a CV, or résumé. The assistant needs the work experience and the intended role. It does not need identity documents, banking details or a referee’s private contact information.
The UK’s National Cyber Security Centre guidance on chatbot risks advises against putting sensitive information into public AI prompts.
What not to share with AI chatbots
Keep these out of ordinary consumer prompts:
- Passwords, recovery codes and one-time login codes.
- Payment-card details and bank account credentials.
- Passport details and government identifiers.
- Private medical information about identifiable people.
- Customer records, confidential contracts and internal business secrets.
Scenario 1: Edit a CV with less personal data
Alex creates a separate working copy and removes contact details, referee information and unnecessary identifying details. Alex then pastes only the experience section that needs editing.
Replacing a name alone is not always enough. A distinctive job title, employer and detailed career history may still identify someone. Remove or generalise details that the task does not require.
Before and after: a safer AI prompt
Before — an oversharing approach:
Rewrite my complete CV. I have attached the original, including my contact details, date of birth and referee information. Make me sound more experienced.
After — a safer, clearer approach:
Improve the wording of the work-experience section below for a customer-support role. Personal details have been removed. Keep all facts accurate, do not invent achievements and mark any point that needs clarification. Return suggested edits for me to review.
The second prompt limits disclosure and sets clearer boundaries. Human checking remains necessary.
Check files and screenshots separately
Before uploading a file, inspect comments, tracked changes, hidden spreadsheet tabs and document properties where relevant. For screenshots, look at browser tabs, notifications, email addresses and background windows.
Use a proper redaction feature for documents and verify the exported copy. Placing a coloured shape over text may leave the underlying text recoverable.
Check the Tool, Privacy Settings and Account Security
Before signing in, Alex checks the tool itself:
- Identify the provider. Reach the app through its official website and check the publisher.
- Assess the task. Can you work with a short extract instead of a file or account connection?
- Read the terms. Find out who receives your data, how it is used and how to delete it.
- Review permissions. Question access to contacts, browser activity, camera or microphone when unnecessary.
A polished design, familiar logo or advertisement does not prove a service is genuine.
Understand AI chatbot privacy controls
Check:
- Whether conversations or uploads may improve models.
- How long information is retained.
- Whether people may review submitted content.
- Whether history saves chats and memory retains details for future use.
- What temporary chats, deletion, export and feedback options actually do.
- Whether connected services receive information.
For example, Google’s Gemini Apps Privacy Hub states that temporary chats are retained for 72 hours but are not used to train its AI models. It also explains that feedback can involve additional processing of associated conversations. These are Gemini-specific terms, not rules for every provider.
Policies vary by provider, subscription, account type and feature. Check the current documentation for the service you actually use. A paid subscription alone does not establish that confidential work data is approved for upload.
Avoid fake apps and protect your login
For account protection:
- Use a unique password stored in a reputable password manager.
- Enable multi-factor authentication where available.
- Protect the email account used for password recovery.
- Keep your device, browser and apps updated.
- Never approve an unexpected sign-in request or share a verification code.
- Sign out after using a shared device.
Multi-factor authentication adds another identity check beyond a password. The NCSC’s password-management guidance supports using distinct passwords and two-step verification.
Reply to Customers Without Oversharing
Alex’s second task is answering a customer who is waiting for an order.
The full email chain includes an address, phone number and unrelated correspondence that the assistant does not need.
Scenario 2: Draft a customer reply
Alex first checks whether the employer permits this use of AI. Using the approved tool, Alex supplies a brief summary without identifying details:
Draft a polite reply to [CUSTOMER]. Their delivery is delayed. We are checking with the courier and will provide an update by [APPROVED TIME]. Do not promise compensation or a delivery date. Keep the reply under 100 words.
Alex then checks the draft against the actual order record and company policy. Only after that does Alex add the customer’s name inside the approved customer-service system.
Australia’s ASD’s ACSC guidance for small businesses recommends clear rules about data uploads, reviewing providers’ data practices and checking AI outputs.
Customer trust also depends on permission. Having access to someone’s information at work does not automatically authorise sending it to another service.
Summarise Work Documents and Limit Permissions
Alex’s third task is summarising a project update containing budget figures, staff comments and externally supplied material.
Scenario 3: Summarise an approved extract.
Alex checks the document’s confidentiality requirements and asks which AI environment is approved for that information.
If approval is unclear, Alex pauses the upload. Instead, Alex can request a generic summary template and fill it in manually.
Where the task is permitted, Alex shares only the necessary extract and requests:
Summarise the supplied extract into decisions, unresolved questions and next actions. Do not add information from outside the extract. Include section references where available and flag anything unclear.
Alex compares the result with the original before circulating it.
Review connected-app permissions
A summarisation task does not automatically require access to an entire drive or mailbox.
Before connecting an app, check whether it can read, change, delete or send information. Prefer access limited to the task, where the service supports it, and remove unused connections.
For Alex, a useful boundary is allowing a draft to be created while keeping sending and file changes under manual control.
Understand prompt injection
Prompt injection happens when malicious instructions in material processed by an AI system try to redirect its behaviour.
Imagine an external project document containing instructions to send unrelated files to a website. Those instructions are part of the document being analysed; they are not legitimate permission from Alex.
The NCSC’s explanation of prompt injection warns that models can confuse untrusted content with instructions, with greater consequences when connected tools are available.
Restrict access, review proposed actions and stop unexpected requests. A sentence telling the assistant to ignore malicious instructions is not a complete security control.
Check AI Answers Before Trusting Them
Alex’s three tasks all require review. The CV must not invent achievements. The customer reply must not create promises. The document summary must not turn uncertainty into a confirmed decision.
AI can produce plausible but incorrect information, often called a hallucination. The NCSC identifies this limitation in its chatbot risk guidance.
Here is a practical way to check AI-generated answers:
- Identify consequential claims. Highlight dates, amounts, names, deadlines and instructions.
- Compare with the original. Check records, documents or official guidance yourself.
- Open cited sources. Confirm they exist and support the specific statement.
- Check context. Information may concern a different country, product version or period.
- Keep uncertainty visible. Remove unsupported claims or clearly mark them as unresolved.
Asking another chatbot can generate ideas for checking, but agreement between chatbots is not independent proof.
For medical, legal or financial decisions, use relevant official information and qualified professional advice. A confident answer should not replace appropriate expertise.
Recognise AI Scams
AI-related risks also arrive outside chatbot windows. A message may impersonate support, promote an imitation app or use synthetic media to make a request persuasive.
The US Federal Trade Commission’s voice-cloning warning explains that scammers can imitate a familiar voice. It recommends independently contacting the person using a number you already know.
Be cautious about secrecy, pressure, unexpected login links and requests for passwords or verification codes. You do not need to prove that a message was AI-generated before treating an unverified request cautiously.
Respond to Accidental Sharing or a Scam
Suppose Alex discovers a private phone number in an uploaded extract. Stop sharing, record what happened and use the provider’s deletion and support options. Exposure does not automatically mean misuse, but it needs attention.
Match your next step to what happened:
| What happened? | Immediate action |
| Password exposed | Change it through the official service and anywhere it was reused; review account activity. |
| Verification code shared | Contact official support and secure the affected account. |
| Unfamiliar app connected | Revoke its access through the connected account’s settings. |
| Bank details shared or money sent | Contact your bank or payment provider promptly through an official channel. |
| Work or customer file uploaded | Notify the responsible manager, IT or privacy team. |
| Public conversation link created | Remove the link and review what it exposed. |
Deleting a visible chat does not necessarily remove every retained copy. Preserve useful evidence without unnecessarily spreading sensitive material.
Where to report
Use the route appropriate to your location and incident:
- US: Report scams through FTC ReportFraud.
- England, Wales and Northern Ireland: Use Report Fraud. Its guidance directs people in Scotland to Police Scotland on 101.
- Australia: Use ReportCyber for cybercrime reporting.
Reporting does not replace immediately securing your account or contacting your bank.
Your 10-Minute AI Safety Review
Set aside ten minutes to review one AI tool, then finish any setup that takes longer:
- I am using the intended, legitimate service.
- My account has a unique password and additional sign-in protection.
- I understand the relevant privacy and retention settings.
- I have permission to use this information with this tool.
- I removed unnecessary personal and confidential details.
- I inspected the complete file, screenshot or text extract.
- Connected apps have only the access needed.
- I will review consequential claims and proposed actions.
- I know whom to contact if information is shared accidentally.
- My family knows how to independently verify an urgent call.
Frequently Asked Questions
Is it safe to put personal information into AI?
Avoid unnecessary personal information. Whether a particular use is appropriate depends on the information, permissions, provider terms and account configuration. Start with the smallest amount of data the task needs.
Does turning off training make a conversation completely private?
No. Training controls do not necessarily eliminate storage, operational processing or security review. Read the provider’s explanation of each setting instead of assuming one switch controls everything.
Can I upload a screenshot instead of typing sensitive text?
A screenshot still shares information. Inspect every visible area and remove unnecessary details before uploading. Changing text into an image does not make its contents confidential.
Are paid AI tools automatically safer than free tools?
No. Compare the actual controls and contractual terms. Some business products have different protections, but paying for an individual subscription does not automatically satisfy workplace requirements.
Can I trust an AI answer if it includes references?
References are a starting point. Open them, check their dates and confirm that they support the claim. A real source can still be cited incorrectly or applied to the wrong situation.
Can AI detect every deepfake or scam?
No detector or visual clue should be your only check. Verify requests for money, access or personal information through an established, separate contact route.
Build a Safer Habit, One Task at a Time
By the end of the day, Alex has a clearer CV, a checked customer reply and a reviewed project summary. Each task has a boundary: limited information, appropriate permissions and human judgement before use.
AI security for beginners is a practical habit you develop with each interaction. Begin with one task today: inspect what you plan to share, check what the tool can access and review what it produces.
Visit AI Security Guides for more practical explanations of AI privacy, online threats and safer technology use.
Educational disclaimer: This article provides general information, not personalised legal, medical, financial or cybersecurity advice. Provider policies change, and no checklist eliminates every risk.