Day 13 of 2160 minShell WalkerHands-on lab
Day 13: Writing Reports That Get Triaged Well
Taught by Mr Elite Updated
By the end of today you can
- Write a bounty report with a title, summary, steps, impact and fix
- Set an honest severity and avoid overclaiming
- Use an assistant to tighten wording without adding invented details
- Audit an assistant's edit and log each change you accept or reject
Warm-up
Answer from memory first, then tap to check. Recalling yesterday is what makes it stick.
Yesterday you met IDOR inside an AI feature. Which two things together produce an IDOR?
A predictable identifier and a missing ownership check on the server.
Why is an allowlist in code a stronger control than a guardrail in the model?
The code enforces the rule every time, while a model guardrail is only a request that can be talked around.
A triager opens your report. They have forty others in the queue and about two minutes for each. If your title says "IDOR bug" and the steps say "just change the ID", they will either ask you three questions or close it as not reproducible. Either way you lose days.
Today you write the report that gets a yes in one read. We will use the IDOR lab as the finding, then ask an assistant to polish the text. Assistants are good at smoothing prose and bad at staying inside the facts, so the second half of today is learning to catch what they add.
What a triager needs from you
A triager is the person at the programme (or the platform) who checks whether your report is real, in scope and worth paying for. They do not know your test account, your browser or your thinking. Everything they need must be on the page.
I use five parts every time, in the same order. Keep each short. The order matters because a triager often decides after the first few lines whether to read on, so the most useful facts go first and the supporting detail follows.
- Title
- One line: bug class, where it is, and the effect. Example: IDOR on /documents/ lets any logged-in user read other users' documents.
- Summary
- Two or three sentences: what is wrong, why it happens, who is affected.
- Steps to reproduce
- A numbered list that someone else can follow cold, with exact URLs, accounts and the result of each step.
- Impact
- What an attacker gains in real terms, stated only as far as your evidence goes.
- Fix
- A short, practical recommendation, such as checking that the document owner matches the session user.
Here is what a good summary looks like for the lab finding. Notice it answers three things: what is wrong, why, and who is affected. It does not mention severity yet, and it does not guess.
The /documents/<id> endpoint returns any document to any
logged-in user. The server checks that the session is valid
but not that the document belongs to that user. Any
registered user can therefore read documents owned by others.Quick check
Which title is best for a triager?
It names the class, the location and the effect, and it claims no severity it has not earned.
Honest severity and the overclaiming trap
Severity means how bad the bug is. Programmes usually rate it from low to critical based on what data or actions are exposed and how easy it is to exploit. Overclaiming means stating more than you proved, such as writing "full database compromise" when you read one document.
Overclaiming costs you trust, and trust is how later reports get read quickly. A triager who catches one inflated claim will check every other line of your report, and every report after it. State what you did, what you saw, and then separate what you proved from what you suspect.
Proved: Logged in as user A, I read document 3, which belongs to user B.
Not tested: Whether documents can be edited or deleted.
Suspected: Other IDs are readable too, based on the sequential pattern.Two things push severity up or down. The first is what the exposed data is worth: a public price list matters less than a private contract. The second is how easy the attack is: needing only a normal account and a changed number is easy. In the lab the data is practice content, so say that its sensitivity is unknown rather than inventing a story about it.
Quick check
You read one other user's document. Which impact statement is honest?
It claims only what you proved and flags the rest as a reasoned suggestion.
Watch me do it: from lab notes to a report
Now the lab, so your report rests on fresh evidence. This runs only against the SecurityElites practice lab, never a real site. I will introduce it in the next section. First, here is how I turn raw notes into the report skeleton, so you see the whole method before you apply it.
mkdir -p ~/day13 && cd ~/day13 && touch notes.md report.md
lsHere is the shape of good notes. They are scruffy on purpose. Notes are for you, so speed and accuracy matter more than style. The values below are examples only, so use what you actually see in the lab.
Account: lab user A
Opened my own doc: /documents/2 -> 200, shows my title
Changed ID to 3: /documents/3 -> 200, doc not in my list
Changed ID to 9999: /documents/9999 -> 404
IDs look sequential (2, 3, 4 all returned documents)# Title
## Summary
## Steps to reproduce
1.
2.
3.
## Impact
Proved:
Not tested:
Suspected:
## Suggested fix
## Evidence
(request, response excerpt, screenshot file name)Notice how the notes above map straight onto the skeleton. The two 200 lines become steps 2 and 3. The 404 tells you the server does respond differently for IDs that do not exist, which is a useful detail for the evidence section. The sequential IDs go under Suspected, not Proved.
Your turn in the IDOR lab
Now the lab itself. We use the DocuShare IDOR lab because you already know the bug class, so your attention goes to the writing. Open it from the lab panel on this page and keep your notes file open next to it. Notes written during the test are far more accurate than notes written from memory.
Then fill in report.md using only what is in notes.md. Aim for steps a stranger could follow in under five minutes. Include what the page should show after each step, such as "the response shows a document that is not listed in my account".
Before you move on, read your finished report aloud once. Anything you stumble over will confuse a triager as well. Fix those spots now.
Using an assistant to tighten, not to invent
Now the assistant. Give it your finished draft and ask for clarity only. Tell it explicitly not to add facts. Use Claude or ChatGPT, whichever you prefer, and check nothing sensitive is in the text first (Day 2 rules still apply).
Edit the bug report below for clarity and concision only.
Rules: do not add any fact, URL, header, version, number or
impact that is not already in my text. Do not change severity.
If something is unclear, ask me instead of guessing.
After the edit, list every change you made as bullet points.
[PASTE report.md HERE]Replace the last line with the full text of report.md, then send it. Copy the edited report from the reply into a new file called report-ai.md in your day13 folder, and keep the assistant's list of changes somewhere handy, since you will compare it with the diff.
Now audit the result. Put the original and the edit side by side and read each sentence. The usual inventions are small and plausible: a made-up header name, "affects all users", a stronger severity word, a claim that data was modified, or a fix that names a framework you never mentioned.
diff report.md report-ai.mdThe output looks like the sample below. The number and letter line, such as 12c12, means line 12 changed. Here the assistant has swapped a modest claim for a broad one, which is exactly what to reject.
12c12
< Any logged-in user can read other users' documents.
---
> This exposes all users' private files.Line 12: REJECTED. "all users' private files" is more than I proved.
Line 20: ACCEPTED. Shorter wording, same meaning.Quick check
The assistant adds the sentence "This exposes all users' private files." You only read one. What do you do?
It is a claim beyond your evidence, and overclaiming damages trust with the triager.
Your turn
IDOR
We use the DocuShare IDOR lab because you have met the bug class already, so your attention goes to the writing. The app only checks that you are logged in, not that you own the document. This is a practice lab built for testing, so it is safe to probe.
What to try
What to notice
- Which part of the request identifies the object, and which part identifies you
- Whether the server ever refuses you, and what that tells you about what it checks
- How little evidence you need to prove the point
If it did not work
Writing steps from memory after the test
Take notes during the test and build the report only from them. If a step is missing, repeat it in the lab.
Letting the assistant polish the severity or impact
Tell it not to change claims, then diff its output and reject anything not backed by your notes.
Vague steps like "change the ID"
Give the exact URL before and after, the account used and the expected result of each step.
Pasting real cookies or personal data into a prompt
Replace secrets with placeholders such as [SESSION_COOKIE] before sharing anything.
Recap
- A triager needs title, summary, steps, impact and fix, in that order
- Reproducible steps matter more than clever wording
- State what you proved, what you did not test and what you suspect
- Overclaiming costs trust and slows later reports
- Assistants smooth prose well but add plausible invented details
- Always diff the edit against your notes and log what you rejected
Cheat sheet
Title formula- Bug class + location + effect
Proved / Not tested / Suspected- Three labels that keep impact honest
diff report.md report-ai.md- Shows exactly what the assistant changed
[SESSION_COOKIE]- Placeholder used in place of real secrets
Edit prompt rule- Clarity only, add no facts, list every change
Before the next day
Mission: Report, edit, audit
Redo the IDOR lab, write a full report from your own notes, have an assistant edit it, and record every change you rejected and why. This builds the habit of treating assistant output as a draft to be checked, which you will rely on in your final reports.
About 40 min
You are done when: A day13 folder containing notes.md, report.md, report-ai.md and changes.md with at least five logged decisions (or all of them, if the assistant made fewer).
Want more? Ask the other assistant to edit the same draft and compare which one invented more details.
Quiz
4 questions. Your score is saved with your progress.
1. Why must steps to reproduce include exact URLs and expected results?
A triager can only confirm what they can repeat, so missing detail turns a real bug into an unconfirmed one.
2. You proved you could read one other user's document. Which severity approach is best?
A reasoned proposal stays within your evidence and lets the programme make the final call.
3. What is the most reliable way to catch invented details in an assistant's edit?
Assistants may deny or miss their own inventions, while a diff plus your notes gives you an independent check.
4. Why tell the assistant to list every change it made?
The list is a second view of the changes. Compare it with the diff, since a change that is not listed is a warning sign.
Finish the day
Mark the day complete to save your progress.
Next: Day 14 covers what happens after you submit: triage feedback, duplicates and how to respond calmly. You will also assemble a personal playbook, and today's report and change log become its first entries.
Your notes
Private to this device. Write what you want to remember from today.