AI Power-Ups — AP Cybersecurity

Where this demo stops and a real attack starts

The "voice shift" above only speeds up or slows down your own recording — it can only ever say exactly what you recorded. A real voice-cloning attack collects samples of a specific person's actual voice and trains a model that generates brand-new sentences in that voice, including things they never said.

This demoSpeed/pitch-shifts audio you chose to record. Can only replay your own words.

Real attackTrained on someone else's voice without consent. Generates new speech they never said.

Impersonation

Clones a voice, face, or writing style from just a few samples, so a message sounds like it's really from someone you trust.

AI-Written Phishing

Writes natural, error-free messages in any tone or language, so the old "bad grammar" red flag stops being reliable.

Data Extraction & Manipulation

Tricks an AI system into revealing information it shouldn't, or feeds it false information so it repeats that misinformation later.

AI-Powered Reconnaissance (OSINT)

Scans public posts and profiles fast, connecting names, routines, and relationships a human would take hours to piece together.

AI-Assisted Malware

Helps write or modify malicious code, or scan for weaknesses, without the attacker needing deep programming skill.

Task A — Phishing Email Blitz

Open your assigned AI chat tool in a new tab. Your goal: get it to produce as many different phishing email variations as you can in 5 minutes, aiming for 20. This is about experiencing the AI-Written Phishing power-up firsthand — speed and scale, not artistry.

Required elements to ask for: urgency, authority impersonation (IT, HR, a manager), a personalization placeholder like [First Name], and a clear call to action (a fake link or a reply request).

Starter prompt — copy it in, then ask for variations:

Write a phishing email pretending to be from a fictional company's IT department. Make it urgent, sound like it's from someone in authority, include a placeholder like [First Name], and end with a call to action like clicking a link or replying with information. Do not use a real link - use http://example-fake-link.test instead. Then generate 5 more variations using different urgency phrases and sign-offs.

Paste the prompt you actually used (feel free to adjust the starter):

0 / 15 words

How many total variations did you get in your 5 minutes? Paste your most convincing output:

0 / 20 words

Paste a second output that felt different from the first:

0 / 20 words

Urgency showed up Authority impersonation showed up Personalization placeholder showed up Clear call to action showed up Sounded professional / error-free

How long did generating this many take, compared to how long it would take to write them all by hand?

0 / 20 words

Were the variations meaningfully different, or mostly superficial rewording? What does that mean for a spam filter trained on old phishing wording?

0 / 20 words

Task B — Urgent Voicemail Script Blitz

Same idea, different power-up: this time you're generating short urgent voicemail-style scripts — the kind an impersonation attack might use. Same 5-minute window, same fictional-only rule.

Required elements to ask for: a claimed identity (who the caller is pretending to be), an urgency/fear trigger, a specific ask (money, info, or an action), and a plausible cover story.

Starter prompt — copy it in, then ask for variations:

Write a short, urgent voicemail script for a fictional scam where the caller pretends to be a family member in trouble and needs money sent quickly. Make it sound panicked and specific, but don't use any real names. Then generate 5 more versions using different fictional emergencies (a car accident, a lost passport while traveling, a mix-up with school security, etc.), each short enough to read aloud in under 20 seconds.

Paste the prompt you actually used:

0 / 15 words

How many total variations did you get in your 5 minutes? Paste your most convincing script:

0 / 20 words

Paste a second script using a different fictional emergency:

0 / 20 words

Claimed identity showed up Urgency / fear trigger showed up Specific ask showed up Plausible cover story showed up

Which fictional emergency felt the most convincing to you, and why?

0 / 20 words

Now that you've seen several versions appear instantly, how would you coach a family member to respond to ANY urgent call like this, regardless of which story it uses?

0 / 20 words

You won't find an "AI-Assisted Malware" version of this task — asking an AI tool to write or modify actual malicious code isn't something this class does, even fictionally. That power-up stays a discussion topic (Scenario 5), not a hands-on one.

Check off anything that's true and easy to find about you online — a real post, a tagged photo, a public profile. Then see what it adds up to.

Your school's name A sport or activity you do Your regular hangout spot An upcoming trip or event A part-time job Your pet's name A close friend's name (tagged in photos) A family member's name

The checklist above is a simulation. These next two are real research tools — one measures how identifiable you are as a person, the other measures how identifiable your browser is as a device. Open each in a new tab and use it for real.

REAL TOOL — Harvard Data Privacy Lab

This tool is built on real research: with just your birth date, gender, and ZIP code — three facts plenty of people post or imply without thinking twice — it estimates how unique that combination makes you among everyone else. Nothing you enter gets saved.

What result did the tool give you for how unique your combination is?

0 / 10 words

Only three data points went into that result. Why does that matter for how easily an AI-OSINT tool could identify someone from public posts alone?

0 / 20 words

REAL TOOL — University of Lille research project

You don't need to give a website your name for it to identify you. Your browser and device quietly hand over a combination of details — screen size, fonts, timezone, installed extensions — that can be unique enough to track without cookies. This tool shows your current fingerprint.

What percentage of browsers did the site say share your exact fingerprint?

0 / 10 words

This tool never asked for your name, birthday, or ZIP code. What does that tell you about how many different ways there are to identify someone online?

0 / 20 words

Scenario 1 — Impersonation

Your grandmother gets a call that sounds exactly like your voice, panicked, saying you're stranded and need money sent to a link right away. She almost sends it — until your mom happens to text you and finds out you were in class the whole time, phone off.

Where would the AI-generated part of this attack have had to happen before the phone call was ever made?

0 / 20 words

What's one thing your grandmother could have done differently to catch this, regardless of how real the voice sounded?

0 / 20 words

Scenario 2 — AI-Written Phishing

A message shows up in the school's shared inbox, written in flawless, professional language, claiming to be from the district's IT office. It asks staff to "verify" their login by clicking a link before the end of the day or lose account access.

What used to make phishing emails easy to spot, and why doesn't that trick work as well anymore?

0 / 20 words

Besides checking for typos, what's a way to verify this message is real?

0 / 20 words

Scenario 3 — Data Extraction & Manipulation

A student finds that phrasing a question a certain way makes a school chatbot repeat a private policy document it was never supposed to share. Separately, someone posts fake "facts" about the school's wifi rules on a public forum, hoping a future AI tool trains on it and repeats the false claim as true.

What are the two different things going wrong in this scenario?

0 / 20 words

Why is it a problem if false information ends up in an AI's training data, even after the original fake post gets deleted?

0 / 20 words

Scenario 4 — AI-Powered Reconnaissance

An attacker spends ten minutes with an AI tool scanning a student's public accounts and their friends' accounts. Within minutes, they know the student's team schedule, favorite hangout spot, and best friend's name well enough to send a message that feels personal and believable.

How is this different from an attacker manually scrolling through profiles for hours?

0 / 20 words

Whose risk went up because of information the student posted — not just the student's own?

0 / 20 words

Scenario 5 — AI-Assisted Malware

A low-skilled attacker with no real coding background uses an AI coding assistant to modify a piece of malware they downloaded, tweaking it just enough that the school's antivirus software no longer recognizes it.

What skill barrier used to stop attackers like this one, and what changed?

0 / 20 words

Does this attacker need to understand how the malware works to use this power-up? Why does that matter?

0 / 20 words

On your own

Verify urgent requests through a channel you already trust — call the person back on a number you already have, not one from the message itself.
Stop trusting "it sounds/reads legitimate" as proof — polish is no longer a reliable signal.
Set a code word with close family for real emergencies.
Assume anything you post publicly could get scraped, and think about the people connected to you, not just yourself.

At an organization

Require independent verification before acting on any financial or account-access request.
Don't rely on authentication that itself depends on a single channel an attacker could fake, like voice.
Keep security training updated for AI-era threats, not just older phishing patterns.
Treat urgency and secrecy ("don't tell anyone else") as red flags on their own.

Pick one defense strategy above. Which scenario from Part 5 would it have stopped, and how?

0 / 25 words