AI security · Practical guide
AI Security Basics for Teachers and Professionals
Start with three questions: what information goes into the tool, what can it access, and what can it do with its output? A writing assistant and an agent that can send emails require different controls.
Classify the information before pasting it
Use public or synthetic examples while learning. Avoid submitting student records, passwords, identity documents, confidential research or internal business material to an unapproved tool. Removing a name may not be enough if other details identify someone.
Use your organisation’s approved service and follow its policy. Check the relevant account’s retention, sharing and training settings; do not assume every personal and enterprise account behaves the same way.
Understand prompt injection
An AI tool may encounter instructions hidden in a document, webpage or message that it was supposed to analyse. For example, a document might tell it to ignore the user and send information elsewhere. This is an instruction from untrusted content, not permission from the user.
Keep external material separate from instructions. Avoid granting unnecessary access, and require human approval for sending, deleting or changing records. A warning in the prompt alone is not a complete defence.
Limit access and protect accounts
Enable multifactor authentication where available, avoid shared logins and review connected apps. Give a tool only the access required for the task. A summarisation task should not need permission to delete files or publish messages.
Treat generated commands and code as untrusted until inspected and tested. Do not paste production secrets into a prompt or embed them in example code.
Worked example: drafting student feedback
Safer workflow: create a fictional response, apply a public rubric, inspect the suggested feedback and adapt it manually. For real submissions, first establish whether your institutional policy and approved account allow that data use.
Riskier workflow: upload a spreadsheet containing names, IDs, marks and personal circumstances to an unknown service, then send its decisions automatically. The problem involves both unnecessary data exposure and insufficient human review.
If something goes wrong
Stop further sharing, preserve relevant details and follow your organisation’s incident-reporting process. If a credential was exposed, revoke or rotate it through the appropriate administrator. Deleting a chat alone should not be assumed to undo disclosure.
Write a short record of what information was involved, which service received it and when it occurred. Avoid copying sensitive information into additional messages while reporting the incident.
Run a ten-minute review
Pick one routine AI task. List its inputs, connections and possible actions. Replace sensitive inputs with synthetic examples for practice. Remove unnecessary access and define the point where a person must approve the result.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
Help me review this hypothetical AI workflow: [describe without secrets or personal data]. Identify information entering the tool, permissions it needs, untrusted content it may read and actions requiring human approval. Suggest a lower-risk practice version using synthetic data.Open the prompt builder →
Pause · Think · Check
Test your understanding
Choose one answer per question, then check your reasoning. This is practice; scores are not saved.