Day 9 of 960 minCipher CadetHands-on lab
Day 9: Capstone: Attack, Report, Defend, and What Next
Taught by Mr Elite Updated
By the end of today you can
- Run a full prompt injection test on the TaskBot practice lab using two techniques from the course
- Write a clear one-page finding with steps, evidence, root cause and fixes
- Review your own report against your Day 8 defence checklist
- Choose three topics to study next in AI security
Warm-up
Answer from memory first, then tap to check. Recalling yesterday is what makes it stick.
Name two layers of defence against prompt injection from Day 8.
Any two of: treating outside text as untrusted, least privilege, human approval for risky actions, filters and monitoring.
Why does least privilege help even when an injection succeeds?
The model can only use the tools and data it was given, so a hijacked model can do less damage.
Imagine you find a weakness in a chatbot and tell nobody useful. You say, "it can be tricked, trust me." The owner shrugs, because nothing in that sentence tells them what to fix. Now imagine you hand them one page: here is what I typed, here is what happened, here is the screenshot, here is why, here is the fix. That page gets acted on.
Today you do the whole job once, start to finish. You attack a practice agent, you write it up, and you check your own work against the checklist you built yesterday. Everything runs inside the SecurityElites lab, which is built for this. Never aim any of it at a real system you do not own or have written permission to test.
The plan for today
A professional test has four parts: attack, evidence, explanation, fix. We have spent eight days learning the pieces. Today we join them. I want you to finish with one document that someone else could read cold and understand.
The lab is called AI Agent Hijacking, and the target is TaskBot. TaskBot is an agent, meaning an AI that can call tools, not just chat. When you ask it to summarise a document, it reads the document and follows instructions inside it as if you had written them. That is indirect injection from Day 4, combined with the tool danger from Day 7.
- Agent
- An AI app that can take actions through tools, such as reading files or sending messages.
- Tool call
- A request from the model to run a function, for example 'read this file'.
- Privileged tool
- A tool the user should not trigger freely, because it touches sensitive data or makes changes.
- Finding
- A written record of one weakness: what, how, why and how to fix it.
Choosing two techniques to combine
The brief asks for at least two techniques from the course. Combining them is not showing off. Real attacks rarely rely on one trick, because each defence blocks something different. Pick a pair that fits TaskBot.
Here are pairings that suit a document-reading agent. Choose one before you touch the lab, and write it in your notes.
- Indirect injection + conflicting instructions
- Plant instructions in the document, and word them so they override the agent's own rules (Days 3 and 4).
- Indirect injection + hidden text
- Plant instructions the human reader would not notice, such as text styled to blend in (Days 4 and 6).
- Indirect injection + tool abuse
- Aim the planted instruction at a privileged tool call (Days 4 and 7).
Quick check
Why is a document the attacker's route in TaskBot, rather than the chat box?
The model cannot reliably separate the user's request from instructions inside content it reads, so planted text gets followed.
Watch me do it: one test, with evidence
I will walk through the method rather than hand you the winning text. The lab gives you the goal: get TaskBot to call a privileged tool through a document. How you word the document is your work. Follow these steps in order.
Attempt 1: direct order at end of doc -> refused, no tool call
Attempt 2: framed as 'note from admin' -> summary only
Attempt 3: + second technique -> tool call firedWriting the one-page finding
Now the part that makes the work useful. A finding is short and plain. A busy developer should understand it in two minutes. Use these six headings in a new file called capstone-report.txt.
TITLE: One line naming the weakness
SUMMARY: Two sentences, what an attacker can do
STEPS: Numbered, so someone can repeat them
EVIDENCE: Screenshot file names and what each shows
ROOT CAUSE: Why it worked
FIXES: From your checklist, in priority order
NEXT: Three topics I will studyFor root cause, be specific. Do not write "the AI is insecure." Write that TaskBot treats document text as instructions and holds a privileged tool the user did not need to reach. That sentence points straight at two fixes: untrusted-content handling and least privilege.
Quick check
Which root cause statement is the most useful to a developer?
It names the exact behaviour and the missing controls, so the fix is obvious.
Review your report against your checklist
Open the ten-item defence checklist you wrote yesterday. Go through your fixes section line by line. For each item, ask whether it would have stopped your attack, and at which step.
That last sentence is what separates a trustworthy report from a boastful one. Defence is layers, never a single wall.
Where to go next
You now have the base. The next areas build directly on it. Pick three for the NEXT heading of your report, and say why each interests you.
- Jailbreaks
- Prompts that talk a model out of its safety rules. Related to Day 3, but aimed at the model's own training rather than the app's instructions.
- Agent security
- Designing agents with scoped permissions, sandboxes and approval steps, building on Days 7 and 8.
- Tool poisoning
- Tampering with the descriptions or outputs of tools an agent trusts, so it misuses them.
- RAG security
- Protecting knowledge sources from planted entries, following Day 5.
Your turn
AI Agent Hijacking
This is the capstone target. TaskBot reads documents and follows instructions inside them, so it combines indirect injection with a tool-using agent. It is the best place to join the skills from Days 3, 4, 6 and 7 in one test.
What to try
What to notice
- Whether TaskBot distinguishes your request from text inside the document.
- What a refusal reveals about the defences in place.
- Which single change finally triggered the tool call.
If it did not work
Changing several things between attempts
Alter one element at a time and log it, so you can explain exactly what worked.
Writing a vague root cause such as 'the AI is unsafe'
Name the behaviour (follows document instructions) and the missing control (no approval, excess privilege).
Taking a screenshot that shows only the reply
Capture the planted text, the response and the tool call or success message together, or use two named screenshots.
Claiming one fix solves everything
List fixes in layers and state what could still get through.
Recap
- A good finding has steps, evidence, root cause and fixes, in plain words
- Combining two techniques mirrors how real attacks beat single defences
- Change one variable per attempt so you can explain what worked
- Root causes should name behaviour and the missing control
- Reviewing your own report against a checklist shows which layers matter most
- Keep every test inside labs or systems you own
Cheat sheet
Attack, evidence, explain, fix- The four parts of every finding.
Indirect injection- Instructions planted in content the model reads, not typed by the user.
Privileged tool- A tool whose misuse causes real harm, so it needs approval and tight scope.
Least privilege- Give the agent only the tools and data it truly needs.
Human approval- A person confirms risky or irreversible actions before they run.
Attempt log- One numbered line per try, with the result.
Before the next day
Mission: Submit your capstone report
Finish the lab, write the one-page report and review it against your Day 8 checklist. This is the proof that you can run a prompt injection test and communicate it clearly. Allow about 45 minutes for the lab and 15 for the report.
About 60 min
You are done when: One page report (capstone-report.txt or a document) with attack steps, an evidence screenshot, root cause, fixes from your checklist and three topics to study next.
Want more? Rewrite the report for two audiences: a two-line summary for a manager and a technical section for a developer.
Quiz
4 questions. Your score is saved with your progress.
1. Why is TaskBot vulnerable to a planted instruction in a document?
The model cannot reliably separate trusted instructions from untrusted content it reads.
2. Why change only one thing between attempts?
A clear cause is needed for an honest root cause section and for repeatable steps.
3. Which fix would most directly reduce the damage even if the injection still succeeded?
Least privilege and approval limit what a hijacked model can actually do.
4. Why should your report admit what could still get through?
No single control is complete, and a realistic report states the remaining risk.
Finish the day
Mark the day complete to save your progress.
What next: There is no Day 10, so your next step is your own study plan. Start with the three topics in your report, most likely agent security, tool poisoning and jailbreaks, and keep practising in labs and systems you own.
Your notes
Private to this device. Write what you want to remember from today.