I85PREP
⌘ K
PSAA study materials
PSAA study resources

PSAA practice questions

One question at a time. Build your PSAA understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Security

In an AI system, which property best distinguishes an adversarial example from ordinary data drift?

  1. The input is deliberately perturbed to cause a specific model error Correct answer.
  2. The input contains missing values that break preprocessing
  3. The model loses accuracy because training labels were noisy
  4. The input distribution slowly changes over many months

Explanation

An adversarial example is an input crafted on purpose, often with a small perturbation invisible to humans, so the model produces an attacker-chosen error. Data drift is an unintentional statistical shift over time, and label noise or missing values are data quality problems rather than deliberate manipulation.

Question 2AI Security

A team wants to reduce the risk that a public-facing chatbot reveals its hidden system prompt. Which control addresses this threat most directly?

  1. Retraining the model on a larger general corpus
  2. Increasing the model temperature during inference
  3. Adding an output filter that blocks verbatim system-prompt content Correct answer.
  4. Enabling request rate limiting per IP address

Explanation

System-prompt leakage is an output-side disclosure risk, so an output filter that detects and blocks verbatim prompt content directly mitigates it. Temperature, extra pretraining data, and rate limiting do not prevent the model from emitting its own instructions.

Question 3AI Security

What is the primary purpose of a model card in AI security governance?

  1. To log every user prompt for forensic review
  2. To document intended use, limitations, and evaluation results Correct answer.
  3. To store the encrypted weights of the trained model
  4. To serve as the runtime API gateway for inference

Explanation

A model card is documentation that records a model's intended use, known limitations, training context, and evaluation results so downstream users can assess risk. It is not an artifact store, a gateway, or a logging system.

Question 4AI Security

An attacker submits a document to a retrieval-augmented generation pipeline that contains hidden instructions telling the model to email internal files to an external address. What is this attack called?

  1. Membership inference
  2. Indirect prompt injection Correct answer.
  3. Gradient masking
  4. Model inversion

Explanation

Indirect prompt injection places malicious instructions in content the model later retrieves, such as a document or web page, rather than in the user's direct message. The model treats the retrieved text as context and may follow the embedded instructions.

Question 5AI Security

Which practice most reduces the risk of sensitive personal data being memorized by a large language model during pretraining?

  1. Deduplicating and scrubbing sensitive records from the training corpus Correct answer.
  2. Quantizing the model to eight-bit weights
  3. Shortening the inference context window
  4. Increasing the number of attention heads

Explanation

Memorization risk is driven by the presence of sensitive records in the training data, so deduplication and scrubbing remove the material a model could reproduce. Architectural changes, context length, and quantization affect performance or efficiency, not what was learned.

Question 6AI Security

A security team needs to verify that a model's responses cannot be altered by a man-in-the-middle on the network between the client and the inference service. Which control applies?

  1. Watermarking of generated text
  2. Differential privacy during training
  3. Input length limits on prompts
  4. TLS with certificate validation for all inference traffic Correct answer.

Explanation

Man-in-the-middle tampering is prevented by encrypting and authenticating the transport channel, which TLS with proper certificate validation provides. Watermarking, differential privacy, and length limits address different threats.

Question 7AI Security

Which activity is an example of red teaming an AI system?

  1. Compiling the model for a new hardware accelerator
  2. Measuring latency under peak production load
  3. Reviewing invoices from the cloud inference vendor
  4. Probing the model with crafted prompts to find policy violations Correct answer.

Explanation

AI red teaming is structured adversarial probing of a system to surface harmful outputs, policy bypasses, and other security weaknesses. Performance testing, compilation, and billing review are unrelated operational tasks.

Question 8AI Security

What does differential privacy add during model training to limit the influence of any single training record?

  1. A second copy of the dataset for validation
  2. An allowlist of trusted users
  3. Calibrated noise in the training or query mechanism Correct answer.
  4. A checksum over the model weights

Explanation

Differential privacy bounds how much any one record can change an output by adding calibrated noise to the training process or to query results. The other options are data handling or integrity measures and do not provide that bound.

Question 9AI Security

An agent powered by a language model can call internal APIs. Which design most directly limits the blast radius if the agent is manipulated through a prompt injection?

  1. Granting the agent a broad service account to simplify operations
  2. Raising the model temperature so responses vary
  3. Caching tool responses to reduce latency
  4. Scoping tool permissions to the minimum actions each task requires Correct answer.

Explanation

Least-privilege tool scoping means a manipulated agent can only perform the narrow actions its current task permits, which limits damage. Broad credentials, higher temperature, and caching do not constrain what the agent is authorized to do.

Question 10AI Security

A model is deployed where attackers can query it repeatedly and observe confidence scores. Why does this increase susceptibility to model extraction?

  1. Confidence scores prevent membership inference
  2. Confidence scores trigger differential privacy automatically
  3. Confidence scores encrypt the model weights
  4. Confidence scores reveal decision boundaries that guide training of a substitute model Correct answer.

Explanation

Rich confidence outputs give an attacker a stronger signal about the model's decision boundaries, which improves the quality of a substitute model trained on the queries. Limiting output detail or rate can reduce this leakage.

Question 11AI Security

An evaluation shows a classifier is robust to small pixel perturbations but fails when the same object appears under unusual lighting in deployment. What does this indicate?

  1. The model has perfect out-of-distribution generalization
  2. The robustness testing covered only a narrow threat model Correct answer.
  3. Differential privacy has been violated
  4. The training set was too small to converge

Explanation

Passing perturbation tests shows robustness only against that specific threat model, not against natural distribution shifts such as lighting changes. Security testing must cover the realistic conditions the system will face.

Question 12AI Security

Which mitigation most effectively reduces training-data poisoning by a small number of malicious contributors to a crowdsourced dataset?

  1. Increasing the learning rate for faster convergence
  2. Data provenance tracking with outlier and influence analysis before training Correct answer.
  3. Removing the output filter on the serving layer
  4. Publishing the model weights openly

Explanation

Provenance tracking identifies where samples came from, and influence or outlier analysis can flag contributions that disproportionately affect model behavior. Together they help remove poisoned samples before training.

Question 13AI Security

A team applies input sanitization that strips known injection phrases, but attackers begin using encoded or paraphrased instructions. What is the key weakness of this approach?

  1. It requires differential privacy to function
  2. It disables TLS for inference requests
  3. It relies on blocklists that cannot enumerate all malicious phrasings Correct answer.
  4. It consumes too much GPU memory during training

Explanation

Phrase blocklists only catch known patterns, and attackers can rephrase or encode instructions to evade them. Robust defenses combine input handling with privilege separation and output controls rather than relying on enumeration.

Question 14AI Security

A membership inference attack against a model trained on patient records succeeds at a high rate. Which factor most likely contributes?

  1. The model is served over HTTPS
  2. The model uses a small context window
  3. The model was quantized after training
  4. The model overfits and its confidence is unusually high on training records Correct answer.

Explanation

Overfitting makes a model behave differently on records it saw during training, and higher confidence on those records gives an attacker a signal for membership inference. Regularization and differential privacy reduce this gap.

Question 15AI Security

Which control best protects a vector database used by a retrieval-augmented generation system from leaking documents a user is not authorized to read?

  1. Storing all embeddings in a single unpartitioned index
  2. Relying on the language model to refuse unauthorized content
  3. Disabling authentication to simplify integration
  4. Filtering retrieved chunks by the requesting user's access permissions Correct answer.

Explanation

Authorization must be enforced at retrieval time so only documents the requester may access enter the context. Relying on the model to refuse is unreliable because the unauthorized text has already been provided.

Question 16AI Security

A safety filter blocks harmful completions, but an attacker splits a prohibited request across several benign-looking turns. Which defense most directly addresses this multi-turn evasion?

  1. Lowering the maximum output token count
  2. Rotating the model's API keys daily
  3. Compressing the model with pruning
  4. Evaluating the full conversation context rather than each message in isolation Correct answer.

Explanation

Multi-turn evasion works because each message looks harmless alone, so the filter must assess the accumulated conversation state to detect the composed intent. Per-message checks miss the pattern.

Question 17AI Security

What is the security rationale for watermarking model-generated content?

  1. It helps identify content as machine-generated for provenance and abuse tracing Correct answer.
  2. It prevents the model from being queried by attackers
  3. It guarantees the model cannot be poisoned
  4. It encrypts the model weights at rest

Explanation

Watermarking embeds a detectable statistical signal so generated content can later be attributed to a model, supporting provenance and abuse investigation. It does not block queries, encrypt weights, or prevent poisoning.

Question 18AI Security

An organization wants to detect when an attacker is systematically probing its model with adversarial inputs in production. Which signal is most useful?

  1. A gradual rise in average response length
  2. A spike in near-duplicate inputs with small perturbations and high query volume Correct answer.
  3. An increase in the model's parameter count
  4. A decrease in the number of active API keys

Explanation

Adversarial probing often appears as many similar inputs with small variations sent at high volume while the attacker searches for a boundary. Monitoring for that pattern supports early detection.

Question 19AI Security

A company fine-tunes a base model on internal data and then shares the fine-tuned weights with a partner. Which risk is most specific to releasing the weights rather than only an API?

  1. The partner loses the ability to query the model
  2. The partner can run extraction and inversion attacks locally without rate limits Correct answer.
  3. The partner cannot measure inference latency
  4. The partner must use TLS to reach the model

Explanation

Releasing weights gives the holder unlimited local access, so extraction, inversion, and probing attacks can proceed without the rate limits or monitoring an API would impose. This is a defining risk of open-weight distribution.

Question 20AI Security

Which approach best prevents an AI coding assistant from executing destructive shell commands suggested by untrusted repository content?

  1. Increasing the assistant's context window
  2. Logging the assistant's suggestions after execution
  3. Fine-tuning the assistant on more code examples
  4. Requiring human approval and sandboxing before any command executes Correct answer.

Explanation

Untrusted content can steer an assistant toward harmful commands, so execution must be gated by human approval and confined to a sandbox that limits damage. Context size, extra fine-tuning, and post-hoc logging do not stop execution.

Question 21AI Security

A red team finds that adding a specific short suffix to any prompt bypasses a model's safety alignment. Which conclusion is most defensible?

  1. Alignment is brittle to a transferable suffix, so input filtering and defense in depth are needed Correct answer.
  2. The model has no safety training at all
  3. The finding is invalid because suffixes cannot affect models
  4. The model's weights have been exfiltrated

Explanation

A short transferable suffix that bypasses alignment shows the safety behavior is brittle rather than absent, so layered defenses such as input filtering and output checks are warranted. It does not prove there was no safety training or that weights leaked.

Question 22AI Security

When evaluating an LLM application for prompt injection, why is testing only the model in isolation insufficient?

  1. Application code cannot affect model behavior
  2. The application's tools, retrieval, and permission checks determine real exploit impact Correct answer.
  3. Injection only affects models without a system prompt
  4. Isolated models always refuse every malicious prompt

Explanation

Real impact depends on what the surrounding application lets the model do, including tool calls, retrieved content, and authorization checks. Testing the model alone misses these paths, so end-to-end system testing is required.

Question 23AI Security

A defender wants a formal bound on how much an attacker can learn about any single training record from model outputs. Which technique provides that guarantee?

  1. Data augmentation alone
  2. Differential privacy with a stated epsilon Correct answer.
  3. Gradient clipping alone
  4. Output watermarking

Explanation

Differential privacy with a quantified epsilon gives a formal bound on how much any single record can influence outputs. Clipping and augmentation are components or heuristics without that end-to-end guarantee, and watermarking is unrelated.

Question 24AI Security

An attacker with only black-box query access wants to craft adversarial examples that transfer to a victim model. Which strategy is most effective?

  1. Train a substitute model on queried input-output pairs and craft examples against it Correct answer.
  2. Query the victim once and stop
  3. Wait for the victim model to be retrained
  4. Add random noise to inputs without any optimization

Explanation

Black-box transfer attacks work by training a substitute model that approximates the victim from observed queries, then optimizing adversarial examples on the substitute so they transfer. Random noise and single queries lack the optimization signal needed.

Question 25AI Security

A team must choose between blocking all retrieved external content and sanitizing it before it enters the model context. Which trade-off analysis is most sound for prompt-injection defense?

  1. Sanitization guarantees immunity to injection
  2. Blocking external content has no effect on utility
  3. Both options are equivalent because retrieval never carries instructions
  4. Blocking is safer but may remove legitimate content, so layered sanitization plus least-privilege tools balances utility and risk Correct answer.

Explanation

Blocking all external content reduces injection risk but can destroy the usefulness of retrieval, while sanitization preserves utility without guaranteeing safety. The sound choice layers sanitization with least-privilege tools and output controls.

Question 26AI Security

In an AI risk assessment, which term describes an attacker deliberately crafting inputs to make a deployed model misclassify or produce attacker-chosen output at inference time?

  1. Federated averaging
  2. Model card publication
  3. Data drift monitoring
  4. Adversarial example attack Correct answer.

Explanation

Adversarial example attacks perturb or design inputs specifically so the model produces an incorrect or attacker-desired output during inference. They target the model's decision boundary rather than its stored data or documentation.

Question 27AI Security

A team wants to reduce the risk that a deployed language model will emit verbatim passages of its proprietary fine-tuning corpus. Which control most directly addresses memorization-based extraction?

  1. Rotating the API keys used to reach the model
  2. Publishing the fine-tuning dataset alongside the model
  3. Increasing the sampling temperature at inference
  4. Applying deduplication and canary-based leakage testing to training data Correct answer.

Explanation

Extraction of verbatim training text is driven by duplicated and memorized records, so deduplicating the corpus and inserting canary strings to test for leakage directly targets the memorization pathway. Canaries let defenders detect whether specific records can be elicited from the model.

Question 28AI Security

An organization runs an LLM agent that can send email and edit calendar entries on behalf of employees. Which design choice best limits the impact of an indirect prompt injection delivered through an incoming email the agent reads?

  1. Relying on the model's system prompt to instruct it to ignore injected instructions
  2. Giving the agent a single long-lived credential with full mailbox and calendar scope
  3. Logging all agent actions to a central SIEM after they complete
  4. Requiring human approval for consequential actions and scoping tools to least privilege Correct answer.

Explanation

Indirect injection works because untrusted content reaches the model inside its context, so the durable defense is to constrain what the agent can do: least-privilege tool scopes plus human confirmation before irreversible actions. Prompt-level instructions are themselves just text in the context and can be overridden by injected content.

Question 29AI Security

A model provider wants to prove to an auditor that a specific suspect model was derived from its proprietary base weights without disclosing those weights. Which mechanism is designed for this?

  1. A checksum of the base model's training configuration file
  2. A TLS certificate pinning the inference endpoint
  3. A salted password hash stored with the base model
  4. A model fingerprint embedded in the base weights that persists through fine-tuning Correct answer.

Explanation

Model fingerprinting or watermarking embeds a detectable signal in the weights so that derivative models can later be identified as descendants, supporting ownership claims without revealing the original weights. The signal must survive fine-tuning to be useful for this purpose.

Question 30Network Security

Which security control is designed to enforce an organization's acceptable-use policy for web traffic by inspecting and filtering HTTP and HTTPS requests?

  1. A stateful packet filter
  2. A load balancer
  3. A hardware security module
  4. A secure web gateway Correct answer.

Explanation

A secure web gateway sits between users and the internet and applies policy to web requests, including URL filtering, malware inspection, and acceptable-use enforcement. Packet filters act on headers and ports rather than web content, while HSMs and load balancers serve unrelated functions.

Question 31Network Security

In a demilitarized zone (DMZ) architecture, which placement is most appropriate for a public-facing web server?

  1. Directly on the internet-facing router interface
  2. In the screened subnet between two firewalls Correct answer.
  3. On the same host as the internal domain controller
  4. On the internal corporate LAN segment

Explanation

A screened subnet (DMZ) isolates internet-facing systems so that a compromise does not give direct access to the internal network. Placing the server on the internal LAN or alongside a domain controller removes that segmentation benefit.

Question 32Network Security

Which protocol provides confidentiality for traffic between a remote user and the corporate network by encapsulating packets inside an encrypted tunnel?

  1. NTP
  2. SNMPv3
  3. ARP
  4. IPsec in tunnel mode Correct answer.

Explanation

IPsec tunnel mode encrypts and encapsulates entire original packets, which is the basis for site-to-site and remote-access VPNs. SNMPv3 and NTP provide management and time synchronization, and ARP resolves local link-layer addresses.

Question 33Network Security

A network engineer needs to detect a port scan that sweeps many destination ports on a single host. Which control type is best suited to this task?

  1. A forward proxy
  2. A VLAN trunk
  3. A signature-based network intrusion detection system Correct answer.
  4. A stateless packet filter

Explanation

A signature-based NIDS inspects traffic patterns and can alert on reconnaissance behavior such as sequential port sweeps. Packet filters enforce rules but do not generate alerts on scan patterns, and proxies and trunks are not detection mechanisms.

Question 34Network Security

Which statement best describes the function of network access control (NAC) in an enterprise LAN?

  1. It replaces the need for authentication servers.
  2. It assigns public IP addresses to internal hosts.
  3. It encrypts all traffic between switches.
  4. It evaluates endpoint posture before granting network access. Correct answer.

Explanation

NAC checks device identity and posture, such as patch level or antivirus status, and then places the device on an appropriate VLAN or quarantine network. It works alongside authentication infrastructure rather than replacing it.

Question 35Network Security

Which record type should be reviewed to determine whether an email domain publishes a policy for rejecting spoofed messages?

  1. A PTR record
  2. A CNAME record
  3. A DNS TXT record containing a DMARC policy Correct answer.
  4. An MX record

Explanation

DMARC policy is published in a DNS TXT record at the _dmarc subdomain and tells receivers how to handle messages failing alignment checks. MX records only identify mail exchangers, and CNAME and PTR records have unrelated roles.

Question 36Network Security

A firewall rule set permits inbound traffic to TCP port 443 from any source. Which security concern does this rule introduce?

  1. It disables encryption on the web server.
  2. It exposes the web service to scanning and exploitation from the internet. Correct answer.
  3. It prevents stateful inspection of return traffic.
  4. It allows unauthenticated access to the management plane.

Explanation

Allowing 443 from any source exposes the service to internet-wide scanning and exploitation attempts, so it should be limited to intended users where possible. The rule does not affect management access, stateful return traffic, or TLS configuration.

Question 37Network Security

Which technology allows multiple isolated Layer 2 networks to share a single physical switch infrastructure?

  1. NAT
  2. DHCP relay
  3. VLAN tagging Correct answer.
  4. Port mirroring

Explanation

VLAN tagging, typically using 802.1Q, lets a physical switch carry multiple logically separated broadcast domains. NAT translates addresses, DHCP relay forwards address requests, and port mirroring copies traffic for monitoring.

Question 38Network Security

An administrator wants to reduce the risk of rogue DHCP servers on an access switch. Which feature should be enabled on user-facing ports?

  1. DHCP snooping Correct answer.
  2. BPDU guard
  3. Port mirroring
  4. Spanning Tree root guard

Explanation

DHCP snooping designates trusted ports for legitimate DHCP servers and blocks offers arriving on untrusted access ports. BPDU guard and root guard protect the spanning-tree topology, and mirroring copies traffic.

Question 39Network Security

A company deploys 802.1X on switch ports. Which component receives the supplicant credentials and forwards them to the authentication server?

  1. The authenticator Correct answer.
  2. The registrar
  3. The supplicant
  4. The certificate authority

Explanation

In 802.1X the authenticator is the switch or access point that relays credentials using RADIUS to the authentication server. The supplicant is the client device, while a registrar and a CA serve unrelated roles.

Question 40Network Security

A security team wants to detect data exfiltration over DNS by observing unusually large TXT responses. Which monitoring approach is most directly applicable?

  1. Reviewing ARP cache entries
  2. Collecting and analyzing DNS query and response logs Correct answer.
  3. Auditing SNMP community strings
  4. Enabling BPDU guard on access ports

Explanation

DNS tunneling and exfiltration appear as anomalous query and response patterns, especially oversized TXT records, which DNS logging makes visible. ARP caches, BPDU guard, and SNMP strings do not reveal DNS payload behavior.

Question 41Network Security

Which characteristic distinguishes a stateful firewall from a stateless packet filter?

  1. It authenticates users with digital certificates.
  2. It maintains a session table to track connection state. Correct answer.
  3. It encrypts traffic between network segments.
  4. It inspects application-layer payloads by default.

Explanation

A stateful firewall records active connections and permits return traffic belonging to established sessions. Application inspection, encryption, and certificate-based user authentication are separate capabilities not implied by statefulness.

Question 42Network Security

A remote-access VPN uses split tunneling. Which risk does this configuration introduce?

  1. The VPN concentrator cannot authenticate users.
  2. All remote traffic bypasses the corporate security stack.
  3. Encrypted tunnels are downgraded to cleartext.
  4. Internet-bound traffic from the client avoids corporate inspection. Correct answer.

Explanation

With split tunneling only traffic destined for corporate networks enters the tunnel, so internet-bound traffic is not inspected by corporate controls. Authentication and tunnel encryption are unaffected by this routing choice.

Question 43Network Security

Which measure best mitigates ARP spoofing on a switched LAN?

  1. Disabling spanning tree on all access ports
  2. Increasing the ARP cache timeout
  3. Enabling dynamic ARP inspection with DHCP snooping bindings Correct answer.
  4. Using static ARP entries on every client for all destinations

Explanation

Dynamic ARP inspection validates ARP packets against trusted DHCP snooping bindings, which blocks forged mappings. Disabling spanning tree creates loops, static entries for all destinations are unmanageable, and longer cache timeouts worsen exposure.

Question 44Network Security

Which control best limits the impact of a compromised internet-facing web server on internal databases?

  1. Placing the database behind an internal firewall that only accepts traffic from the web tier Correct answer.
  2. Publishing the database server in the DMZ
  3. Enabling port mirroring on the core switch
  4. Configuring a longer DHCP lease time

Explanation

Segmenting the database behind an internal firewall with least-privilege rules limits lateral movement from a compromised web tier. Mirroring, lease duration, and DMZ placement do not restrict database reachability.

Question 45Network Security

Which protocol should be used instead of Telnet to manage network devices securely?

  1. FTP
  2. HTTP
  3. SSH Correct answer.
  4. TFTP

Explanation

SSH encrypts management sessions and provides authentication, whereas Telnet sends credentials in cleartext. TFTP, HTTP, and FTP all lack the encrypted interactive session needed for secure device administration.

Question 46Network Security

A SIEM correlates firewall denies, IDS alerts, and authentication failures from a single source IP within minutes. Which activity does this pattern most likely indicate?

  1. A routine backup window
  2. An automated attack or scanning campaign Correct answer.
  3. A spanning-tree topology change
  4. A DHCP lease renewal

Explanation

Rapid, correlated failures across multiple controls from one source are characteristic of automated scanning or attack tooling. Backup activity, topology changes, and lease renewals do not produce this multi-control failure signature.

Question 47Network Security

Which technique provides the strongest protection against on-path attackers reading credentials during remote administration?

  1. Reusing a shared administrator password across devices
  2. Requiring mutual TLS or host-key verification on all management sessions Correct answer.
  3. Disabling logging to reduce overhead
  4. Relying on network address translation

Explanation

Mutual TLS or strict host-key verification authenticates both ends and prevents an on-path attacker from impersonating a management endpoint. Shared passwords, NAT, and reduced logging do not stop interception.

Question 48Network Security

An organization wants to enforce that only approved applications reach the internet from user endpoints. Which architecture most directly supports this?

  1. A DHCP relay agent
  2. A forward proxy with allow-listing and TLS inspection Correct answer.
  3. A spanning-tree root guard
  4. A reverse proxy in front of internal servers

Explanation

A forward proxy mediates outbound client requests and can allow-list approved destinations, with TLS inspection where policy permits. A reverse proxy protects inbound services, and the other options address switching and addressing.

Question 49Network Security

Which statement about IPsec ESP compared with AH is accurate?

  1. ESP operates only in transport mode.
  2. Both protocols provide identical services.
  3. AH encrypts payloads, while ESP authenticates headers only.
  4. ESP provides confidentiality and integrity, while AH provides integrity and authentication without confidentiality. Correct answer.

Explanation

ESP encrypts the payload and offers integrity and authentication, whereas AH provides integrity and origin authentication without encryption. ESP can operate in either transport or tunnel mode.

Question 50Network Security

An attacker on the same wireless network captures handshake traffic to crack a pre-shared key offline. Which configuration most reduces this risk?

  1. Reducing transmit power
  2. Using WPA3-SAE or 802.1X/EAP instead of a shared WPA2-PSK Correct answer.
  3. Hiding the SSID broadcast
  4. Enabling MAC address filtering

Explanation

WPA3-SAE and 802.1X/EAP resist offline dictionary attacks against captured handshakes, unlike WPA2-PSK. Hiding the SSID, MAC filtering, and power reduction are weak obscurity measures.