What Else Came Along?

Subtitle: Inspect what a harmless-looking file can carry before deciding whether a tool should receive it.

What you will make

A File Safety Note that separates visible content, non-obvious content, generated observations, and the next verification step. You will inspect an RS-provided synthetic DOCX pair; no personal or workplace file is needed.

Lab materials

• 📎 Resource · What Else Came Along Training — Visible workshop notes plus harmless hidden training details.
• 📎 Resource · What Else Came Along Sanitized — The same visible notes with the targeted marker and fictional properties removed; the reveal guide documents what remains.
• 📎 Resource · Fixture reveal guide — The delayed reveal and calibration route for learners who do not upload the synthetic file pair.
• 📓 Notebook · Lab Notebook — An optional place to record the work and download a local copy.
• 📎 Resource · Facilitator guide

Before you start

Use the supplied files above rather than a personal file. The training copy contains benign visible notes, one harmless white-on-white marker, and fictional document properties. The sanitized copy keeps the visible notes and removes those targeted training details. “Sanitized” describes that bounded procedure; it does not mean the package is free of every non-visible detail.

AI is optional. A facilitator reveal, the 📎 Resource · fixture reveal guide, or a partner comparison meets the outcome. If a tool is used, upload only the supplied training artifact. Do not use a real personal, family, school, client, or workplace file.

No recording is required.

The core idea

A file is a package, not only the page you see. It may carry properties, comments, revisions, hidden or low-contrast text, custom data, relationships, or instructions that another system can parse. Inspection reduces uncertainty; it does not prove universal safety. One model response never proves that a file or tool is safe.

Use four practical choices: share / sanitize / substitute / stop. Limit any connected tool to the smallest information and authority needed.

Inspect the supplied file pair

Complete the activity here. Open the optional 📓 Notebook · Lab Notebook beside this page when you want a place to record your work.

1. 👀 Read only what is visible. Open the training copy and record only what you can directly observe on the page. Do not infer that the file contains nothing else.

2. 🔮 Predict what else may travel. Before revealing anything, list possible document properties, comments, revisions, hidden text, file-format or tool metadata, links, package relationships, embedded objects, or instructions. Mark each as a prediction, not a finding. A reasonable prediction remains useful even when this fixture does not contain it.

3. 🧩 Reveal the supplied training details. Use only the Step 3 section of the 📎 Resource · fixture reveal guide. Confirm the white-on-white marker RS-SYNTHETIC-MARKER-ORANGE-LANTERN-271 and the fictional Author, Company, and Subject properties. These are deliberately harmless training details. Keep your other predictions open until the comparison.

4. 🧼 Compare the sanitized copy. Confirm that the visible workshop notes remain while the targeted marker and fictional properties are absent. Then continue to the Step 4 section of the 📎 Resource · fixture reveal guide: inspect the documented generator metadata that remains in both packages and calibrate the predictions that did not appear. Name what the procedure preserved, removed, and left behind.

5. 🚦 Choose share, sanitize, substitute, or stop. For the supplied practice file, explain which choice fits and why. Then describe how your choice would change if the file belonged to another person or contained client information.

6. 🤖 Optionally ask a bounded tool. If you choose an AI route, upload only the supplied training artifact and ask what content it can identify. Give it no external actions, email, storage, or connected-account authority. Record whether you supplied a file, extracted text, or human notes.

7. 🔎 Separate result from proof. Label direct observations, tool-generated claims, and unanswered questions. Choose one independent check—for example, package inspection, document properties, a second tool, or a human review—before deciding what to do next.

Why this process works

Prediction makes hidden assumptions visible. The supplied pair lets learners compare targeted changes without risking real information. Calibrating absent predictions prevents careful inspection from being mistaken for overreaction. A bounded optional upload demonstrates that a model may notice some content and miss other content; the learning target is the disclosure decision, not a tool contest.

This fixture demonstrates hidden carriage, not prompt injection: its marker is inert text and does not tell a tool to act. OWASP describes the separate risk of indirect prompt injection, in which external content influences model behavior, and recommends separating untrusted content, limiting privileges, and keeping human approval around higher-risk actions. Microsoft documents several kinds of hidden document data and notes that white-on-white text may evade the ordinary hidden-text inspector. See the 📎 Resource · source note.

Check your result

• I separated visible content from predicted or revealed content.
• I treated the marker and metadata as synthetic training details.
• I compared what the sanitized copy preserved and removed.
• I named the tool input and authority boundary, if I used one.
• I did not treat one response as proof of safety.

You are done when

• A prediction made before the reveal
• A visible-content observation
• The revealed marker and fictional metadata
• A comparison with the sanitized copy
• A share, sanitize, substitute, or stop decision
• One independent verification or next step

Try another version

• Repeat from the on-page reveal without downloading either file.
• Ask a partner to predict a different kind of hidden content.
• Compare package inspection with a tool-generated description.
• Write a five-line safety pause for a future file task.