Bypassing Defensive Controls GOAA Practice Questions (Page 8)

Questions 36–40

1. 36application · medium

   A red team is testing an email gateway that uses a proprietary NLP model to filter phishing. The team has no access to the model's parameters or gradients, but they can submit emails and observe whether each is blocked. They have a local surrogate model trained on similar data. Which technique would allow them to craft emails that evade the gateway?

   • ADirectly compute the gradient of the gateway's loss function using backpropagation.
   • BSend a large volume of benign emails to train a local copy of the gateway.
   • CUse a dictionary attack to find words that are not in the gateway's vocabulary.
   • DUse the surrogate model to generate adversarial examples and test them against the gateway, iterating based on the gateway's responses.
   Reveal Solution Discussion Select an answer first
2. 37application · medium

   A security team deploys an AI-based malware detector that uses a deep neural network to classify executable files. During a red-team exercise, an analyst wants to craft a malicious executable that evades this detector without having access to the model's internal parameters. The analyst only has the ability to submit files and observe whether they are flagged. Which approach is most appropriate?

   • AExploit a known vulnerability in the detector's input parser to cause a crash, bypassing detection entirely.
   • BUse a gradient-based attack directly on the target detector by computing the gradient of its loss function with respect to the input.
   • CSubmit random mutations of the malicious file until one is not flagged, without any systematic strategy.
   • DTrain a surrogate model on a public malware dataset and use its gradients to craft adversarial examples, then test them against the target detector.
   Reveal Solution Discussion Select an answer first
4. 38expert · hard

   A company uses an AI-based email security gateway that employs both a spam filter and a phishing detector. The phishing detector is a deep learning model that analyzes email text and URLs. A penetration tester has successfully evaded the phishing detector by using homoglyphs in URLs. The security team wants to mitigate this vulnerability without blocking legitimate emails that use internationalized domain names. Which mitigation is most appropriate?

   • AUse a blacklist of known homoglyph characters.
   • BIncrease the phishing detection threshold to require higher confidence.
   • CBlock all emails containing internationalized domain names.
   • DNormalize URLs by converting homoglyphs to their ASCII equivalents before feeding them to the model.
   Reveal Solution Discussion Select an answer first
5. 39application · medium

   A security analyst is testing an AI-based malware classifier that is deployed as a black box. The analyst wants to generate adversarial examples that are misclassified by the classifier. The analyst has a large dataset of benign and malicious files. Which approach is most likely to succeed?

   • ADirectly compute the gradient of the black-box classifier using backpropagation.
   • BSubmit the same malicious file multiple times to see if the classifier's output varies.
   • CUse a random search to modify file bytes until the classifier changes its prediction.
   • DTrain a surrogate classifier on the same type of data and use its gradients to craft adversarial examples, then test them on the black-box classifier.
   Reveal Solution Discussion Select an answer first
6. 40application · medium

   An organization uses an AI-based web application firewall (WAF) that detects SQL injection attacks using a machine learning model. A red team finds that adding certain characters to the payload evades detection. The organization wants to implement a defense that is robust against such evasion. Which approach is most effective?

   • ADeploy the WAF on a more powerful server to handle more traffic.
   • BUse adversarial training by including adversarial SQL injection examples in the model's training set.
   • CSwitch to a signature-based WAF that uses known attack patterns.
   • DIncrease the WAF's logging verbosity to capture more details about the attacks.
   Reveal Solution Discussion Select an answer first