Scroll to turn the page
Cybersecurity · AP Cybersecurity
social engineering, passwords, public Wi-Fi and AI
An adversary rarely breaks a system head-on. It finds the easiest way in: a person who can be talked into acting, a password that can be guessed, a network that can be faked.
The figure runs three such avenues against one target, one after another. A message built to make the target click (social engineering); a stream of guesses at a weak password (weak authentication); and a fake Wi-Fi network with nearly the café’s name, which passes the target’s traffic to the adversary (public Wi-Fi).
This chapter is Unit 1 of AP Cybersecurity, about 10 class periods in the course framework. It covers social engineering (1.1), signs of a password attack and stronger authentication (1.2), adversaries and attacks on public networks (1.3), how adversaries use AI (1.4), and how defenders use it (1.5).
The unit starts the course’s three skills: explaining how a vulnerability creates risk, finding ways to mitigate it, and detecting attacks from signs of suspicious activity in logs, called indicators of compromise (IoCs). The exam weights skills, not units: Analyze Risk, Mitigate Risk and Detect Attacks are each 25–40% of the multiple-choice section.
manipulating people, not machines
Social engineering attacks use psychological tactics to manipulate users into revealing sensitive information, downloading a malicious file, or clicking on a malicious link.
Getting a target to reveal information is called elicitation. The other two actions hand the adversary a way onto the target’s device or to a page where the target types a password.
Social engineering can be performed in person, but it is often done by email, by text message or through social media messages. The figure sends one message down each channel and shows the action it is after: by email, reveal sensitive information (elicitation); by text message, download a malicious file; by social media message, click a malicious link.
No technical flaw is needed: the attack works if the target does what the message asks. That is why the defense starts with recognizing the message.
why people act before they think
Adversaries often use two psychological tactics. Intimidation is threatening a target with negative consequences if they don’t comply. Urgency is creating reasons why a target should act quickly.
Both rely on common psychological principles. Intimidation leverages a natural human aversion to negative consequences: by drawing attention to them, the adversary uses fear to incite the target to act. Urgency leverages the natural response to react quickly to time-sensitive needs: a target who feels pressured to respond doesn’t take the time to consider whether the action is reasonable or safe.
The figure lights the cues of a hypothetical phishing email in turn:
From: PaperTrail Security <alerts@papertrai1-support.example> To: r.diaz@school.example Subject: [URGENT] Account deleted in 2 hours We detected a sign-in from a new device. Verify your identity within 2 hours, or your account will be permanently deleted and all of your files will be lost. Verify now: papertrai1-support.example/verify — The PaperTrail Account Team
Urgency: “within 2 hours”, in the subject and again in the body. Intimidation: the account “permanently deleted”, the files “lost”. And two more indicators: the sender’s domain papertrai1-support.example imitates the real papertrail.example with a one for the l, and the link goes to that look-alike domain.
The safe response is to slow down: don’t use the message’s link; go to the real site or app yourself and check.
information, codes, and control of a device
The damage of a social engineering attack depends on what the victim hands over.
Personal information that could lead to impersonation: name, phone number, address, workplace, pets’ names, birthdate. These are the same facts many websites use as challenge questions to verify identity. In the figure the victim’s pet’s name, birthdate, school reach the adversary, who answers “Name of your first pet?” with “Biscuit”.
Secure information like a one-time password (OTP) or login code. With the code 482 913, the adversary can log in to a service as the victim.
Malware. A download or a click can install malware on the device, steal information from the web browser (here saved passwords and browser cookies), or direct the victim to a website where their login credentials are captured.
reading an account’s login log
In an online password attack, adversaries try logging in to a device or service using common passwords, common password patterns, or stolen passwords.
Signs of one, in a log of logins: many failed attempts over a short duration, login attempts at unusual times, and attempts from unknown devices.
A hypothetical account’s activity:
day time IP address device result Sat 10:43:02 198.51.100.23 known success Sun 20:51:17 198.51.100.23 known success Tue 03:12:04 203.0.113.77 unknown FAILED Tue 03:12:08 203.0.113.77 unknown FAILED Tue 03:12:12 203.0.113.77 unknown FAILED Tue 03:12:16 203.0.113.77 unknown FAILED Tue 03:12:20 203.0.113.77 unknown FAILED Tue 03:12:24 203.0.113.77 unknown FAILED Tue 03:12:28 203.0.113.77 unknown FAILED Tue 03:12:32 203.0.113.77 unknown FAILED Tue 03:12:36 203.0.113.77 unknown success Tue 18:34:50 198.51.100.23 known success
The owner logs in from one known laptop at 198.51.100.23, on weekends and in the evening (blue on the graph). Then, early on a Tuesday, 8 failed attempts come from an unknown device at 203.0.113.77, 4 s apart, 28 s from the first to the last: every sign at once (red). The success that follows is the worst line in the log: the attack worked, and the account is no longer the owner’s alone.
A burst this regular is a script, not a person typing. Counting failures per IP address and per minute is how a defender spots it.
patterns, and dictionaries built from a life
Many people create passwords with common patterns, and adversaries know them.
The patterns: start with one or two words, add a two-digit number (often a year), and put a special character at the end; include the names of family or pets; include personally significant dates.
So an adversary builds a dictionary of possible passwords from personal information gathered about the target (birthday, anniversary, names of pets and family), and an automated tool submits them one after another.
In the figure the adversary knows a hypothetical target’s pet (Biscuit), sister (Maya) and birth month and year (3/2009). 2 words × 2 numbers (09, 03) × 3 symbols (! ? #) gives 12 candidates. The target’s password, Maya09!, follows the pattern exactly: it is guess 7.
passwords a dictionary won’t find
Users should create passwords that are long, random and unique, and avoid names, dates, or other personally meaningful words or numbers.
Long and random make guessing hopeless. The figure counts the passwords an attacker must try (log scale), at a hypothetical 1,000 guesses a second; on average the right one turns up halfway:
| password | possible | average time |
|---|---|---|
word + two digits + symbolMaya09! | 10,000 common words × 100 two-digit numbers × 32 symbols = 3.2 × 107 | 4.4 hours |
five random wordsgravel-lantern-orbit-maple-sleet | 7,7765: five words, each from a list of 7,776 = 2.8 × 1019 | 4.5 × 108 years |
16 random charactersq7#Lm2$vR9!xT4@k | 9416: 16 characters, each any of 94 on a keyboard = 3.7 × 1031 | 5.9 × 1020 years |
The pattern falls in 4.4 hours, and a dictionary of the target’s own words is far smaller still. Five random words take about 4.5 × 108 years.
Unique matters because stolen passwords get tried on other sites: a password reused from a breached site is not a guess at all.
A password manager can generate and store strong passwords (like the 16 random characters), or a user may create long, unique passphrases for their accounts.
a password is not enough on its own
When available, users should enable multifactor authentication (MFA), which requires extra proof of identity, such as a one-time code, in addition to the password.
MFA is an extra layer of security: a guessed, phished or stolen password alone no longer opens the account.
The figure runs one login form twice. An adversary with the stolen password gets past the first step, then is asked for the code sent to the owner’s phone, which the adversary doesn’t have: denied. The owner types the same password, then the code (731 806, hypothetical) from the phone: signed in.
MFA is one reason the 1.1 attacks ask for one-time codes. A victim who reads the code out to an adversary gives away the second factor too: a real service never asks for it.
skill and motivation
Adversaries can be classified by their skill levels.
Low-skilled adversaries rely on malicious cyber tools created by others that can be purchased online. The tools they use exploit known vulnerabilities.
High-skilled adversaries can create new malicious cyber tools or modify existing ones to adapt to new defensive techniques and tools. They can also discover undocumented vulnerabilities, known as zero days.
In the figure the low-skilled adversary buys a tool and uses it on one of the system’s 3 known vulnerabilities; the high-skilled adversary builds one and finds a vulnerability nobody had listed.
Adversaries also have a variety of motivations: greed, recognition, a cause, revenge, politics and beliefs. Who the adversary is, and why, tells a defender what they are likely to attack and how hard.
a look-alike wireless network
In an evil twin attack, an adversary sets up their own wireless access point (WAP) with a service set identifier (SSID), the network’s name, similar or identical to a target network.
Victims may unknowingly select and connect to the evil twin, allowing the adversary to capture their network traffic. In the figure a café’s network (hypothetical) is “Harbor Books Wi-Fi”; the adversary’s, beside it, is “Harbor Books WiFi”. The laptop joins the twin, and every packet passes through the adversary’s hands.
What the adversary can read depends on the protocol. It cannot read traffic that uses an encrypted protocol like HTTPS: those packets stay sealed. An unencrypted DNS query (the laptop asking where mybank.example is) can be read, and shows which sites the victim visits.
The defense: verify that the name of any wireless network you join exactly matches the name of the network you intend to join. A fake network can also show a fake login page (a captive portal) asking for an account’s password; don’t give it one.
two more wireless attacks
In a jamming attack, an adversary floods an area with a strong electromagnetic (EM) signal in the same frequency range as the wireless network.
The jamming signal prevents legitimate traffic between the access point (AP) and its users. An attack that prevents users from accessing resources is a denial of service (DoS) attack. On the left of the figure, while the jammer is on, the users’ traffic stops.
In a war driving attack, adversaries try to detect wireless network beacons, the announcements an access point sends out, while driving or walking around a target.
A detected signal tells the adversary the type of wireless network used (here the hypothetical “Northfield-Staff”, WPA2) and shows where the signal extends outside the physical building. On the right, the car hears the staff network on about 25% of its lap, all of it outside the walls: places an adversary could work from without ever going in.
encryption, and a VPN
Most internet protocols are encrypted to protect network traffic. But not everything is, so individuals may weigh the sensitivity of their data in choosing whether to join an unencrypted Wi-Fi network at all.
Some data, like DNS queries, can be vulnerable on an unencrypted network: anyone nearby, or the network’s owner, could learn which sites are being visited.
Individuals may consider a virtual private network (VPN), which encrypts all their traffic to the VPN operator’s system. In the figure the traffic runs inside the tunnel past the café’s access point and the service provider, which see only encrypted data.
The VPN does not make the traffic invisible to everyone. It is decrypted at the VPN server: the VPN provider can view the traffic. A VPN moves the trust from the café and the service provider to the VPN operator.
Together, the habits are: join only the network whose name exactly matches the one you mean, prefer encrypted connections such as HTTPS, and use a VPN when the data matters.
impersonation, and a shared secret
Adversaries can use AI-powered tools that leverage existing voice and image samples of a person to create a digital avatar of that person.
With one, an adversary can impersonate someone over the phone or even on a video call, which can lead to financial loss or the sharing of sensitive or private information. As more organizations adopt voice-based authentication, voice impersonation has a larger potential impact.
In the figure (hypothetical) 45 s of a teenager’s voice, from 3 public video posts, is enough for a tool to clone it. The cloned voice calls a relative, frantic, asking for bail money.
The protection: establish shared secrets with close friends and relatives, a secret word or phrase known only to the two of you, to authenticate each other in high-stakes situations. The relative asks for the word, and the clone has no answer.
And enable MFA: if an adversary clones a voice to get into a system with voice authentication, requiring a second factor can still keep them out.
phishing, leaks, poisoning and recon
Generative AI tools, like large language models (LLMs), help adversaries in several ways.
Phishing in any language. Unnatural language has been a way to spot phishing written by non-native speakers. With an LLM, messages in any target language read as though a native speaker wrote them, so that cue is gone.
Extracting information. Adversaries can craft prompts that extract secure or sensitive information from an LLM; it can come from user input and from the large data sets used to train the model. Poisoning. Adversaries can publish or modify websites with false information so that it is included in an LLM’s training set and the model repeats it (in the figure, a fake help-line number).
Reconnaissance and code. AI tools can scan the internet to gather what a target has posted on social media and public websites, and AI coding tools can help write new malware, modify application code to act maliciously, or find vulnerabilities in large code bases.
Protections: don’t enter personal or sensitive data into AI tools such as chatbots or virtual assistants (some feed user input back into training), and evaluate AI output carefully, verifying it with reputable, stable, non-AI-based sources.
configurations, code and detection rules
Defenders use AI-powered tools too, and the rule is the same each time: a knowledgeable person reviews the recommendation before it is used.
AI tools can review security configurations, like firewall rules and access controls, and recommend more secure options: checked by a knowledgeable security technician. They can analyze application code to identify vulnerabilities and recommend mitigations: reviewed by a knowledgeable programmer. They can suggest rules for automated detection systems: reviewed by a knowledgeable detection engineer before being added.
In the figure (hypothetical) a tool makes 4 recommendations:
| recommendation | reviewed by | result |
|---|---|---|
| firewall rule: block inbound telnet (port 23) | security technician | approved |
| code fix: validate input before the database query | programmer | approved |
| detection rule: alert on 10 failed logins in 1 min from one IP | detection engineer | approved |
| access control: remove every account’s access to the file server | security technician | rejected |
3 are sound and go to the live system. The last would cut every user off from the file server, so the technician rejects it. The tool is fast; the reviewer knows what the system is for.
millions of events, a few alerts
Of the millions of digital events that happen on a network every day, some likely represent an adversary at work. Humans cannot carefully examine them all.
A hypothetical network logs 4,000,000 events a day. An analyst who looks carefully at one a minute covers 480 in an 8-hour shift: 8,333 shifts for one day’s events.
AI-powered tools can be trained to quickly analyze digital events and sort those that are likely malicious from those that are harmless. If the tool flags 0.01% of them, that is 400 a day: one analyst’s shift.
The tools can be programmed to alert human cybersecurity personnel when likely malicious activity is detected, or to take specific corrective actions based on its type: in the figure, blocking the address 203.0.113.77.
So threat-detection and response teams catch malicious activity and intervene quickly, preventing loss, harm, damage and destruction to digital infrastructure and data.