Safety: Share Less, Notice More

Audience

This module is for beginning learners, families, educators, facilitators, and small teams who want a calm way to decide what should and should not enter an AI-supported activity.

No AI account, upload, personal information, or direct tool use is required. Learners may read, discuss, inspect the supplied synthetic files, observe a facilitator demonstration, or use an offline comparison.

Purpose

Help learners make practical, non-fearful choices before text, files, images, links, or tool access enter an AI-supported activity.

Four Choices

• Share: the information is necessary, appropriate, low-risk, and yours to share.
• Sanitize: remove names, identifiers, hidden content, metadata, or unnecessary detail while preserving what the task needs. Inspection reduces uncertainty; it is not a guarantee that nothing else will travel with a file.
• Substitute: use a synthetic, fictional, or facilitator-provided example that exercises the same skill.
• Stop: do not proceed when ownership, consent, consequence, or tool authority is unclear.

Worked Example

Suppose a parent wants help rewriting a school email that names a teacher and another child.

• Share: not as written; the message contains information about people who did not choose to share it.
• Sanitize: replace names with roles, remove the forwarded thread, and check the file or pasted text for details the task does not need.
• Substitute: practice with a fictional parent, teacher, and student facing the same communication problem.
• Stop: do not proceed if the email includes a third party's private message or the consequences and consent remain unclear.

The right choice can change as more becomes visible. Sanitizing reduces what is exposed; it does not transfer responsibility to the tool.

Learning Outcomes

Learners can:

• distinguish their information from private or third-party information;
• notice that a file can carry comments, revision history, metadata, hidden text, or other content beyond the visible page;
• explain indirect prompt injection as untrusted instructions arriving through external material such as a file or webpage;
• limit a tool to the smallest information and authority needed;
• choose share, sanitize, substitute, or stop;
• explain why one model response does not prove that a file, service, or connected tool is safe.

Practical Safety Pause

Before using a tool, ask:

• Whose information is this?
• What else may travel with it?
• What is the smallest useful version?
• What can the tool read, retain, retrieve, or act on?
• Who checks the result and who can stop the action?

Related Lab

• 🧪 Lab · What Else Came Along? — a Beta practice for inspecting a supplied synthetic DOCX, revealing harmless hidden content and fictional metadata, and comparing a sanitized copy without requiring an upload.

Current Guidance Boundary

OWASP describes indirect prompt injection as external content influencing a model in unintended ways and recommends separating untrusted content, constraining privileges, and requiring human approval for higher-risk actions. Microsoft documents that Office files can contain properties, personal information, revisions, hidden text, and other non-obvious material; it also notes that white-on-white text is not detected as ordinary hidden text by Document Inspector. These sources support the safety pause, not a promise that inspection makes every file safe.

Sources

• OWASP GenAI Security Project: LLM01 Prompt Injection — accessed 2026-08-17.
• Microsoft Support: Remove hidden data and personal information by inspecting files — accessed 2026-08-17.

Tool Boundary

AI is optional. Learners may use supplied fixtures, screenshots, a facilitator reveal, or an offline comparison. If a tool is used, give it only the information and authority required for the bounded practice. A model response does not prove that a file, service, or connected tool is safe.

Privacy Boundary

Use only the supplied synthetic fixtures or another approved fictional example. Do not use real personal, family, school, client, workplace, health, legal, or financial files. No recording, upload, or disclosure is required.