Bypassing Defensive Controls GOAA Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A security team deploys a computer-vision model to detect weapons in luggage X-ray scans. During a red-team exercise, the team wants to test whether an attacker could conceal a small knife by placing a printed sticker on the blade. The team has white-box access to the model and can compute gradients. Which approach would most directly produce an evasion that survives physical printing?

   • AUse a random noise overlay on the knife image and print it, hoping the detector fails due to added texture.
   • BAdd a small, imperceptible perturbation computed via the fast gradient sign method (FGSM) to the knife image, then print the modified image on a sticker.
   • CApply a Gaussian blur to the entire knife image and print it to reduce edge detection.
   • DCrop the knife image to remove the blade from the X-ray frame before printing.
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A company uses an AI-based spam filter that has been evaded by adversarial emails. The security team wants to evaluate the effectiveness of different countermeasures. They have a dataset of known adversarial examples and want to measure how well each countermeasure reduces the evasion rate. Which evaluation metric is most appropriate?

   • AThe time taken to classify each email.
   • BThe number of false positives on legitimate emails.
   • CThe model's accuracy on a clean dataset of legitimate emails.
   • DThe percentage of adversarial examples that are correctly classified as spam after applying the countermeasure.
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A penetration tester is attempting to bypass an AI-based malware detection system that uses a deep learning model. The tester has limited query access to the system and wants to minimize the number of queries. Which attack technique is most suitable?

   • AUse a white-box attack by assuming access to the model's parameters.
   • BUse a transfer attack by crafting adversarial examples on a surrogate model and testing them on the target.
   • CUse a query-based attack that requires many queries to estimate gradients.
   • DUse a brute-force approach by trying many random mutations.
   Reveal Solution Discussion Select an answer first
5. 34foundation · easy

   What is the primary goal of a red-team exercise targeting AI-driven defensive controls?

   • ATo train the AI model on additional benign data
   • BTo replace the AI defenses with traditional rule-based systems
   • CTo demonstrate that AI defenses are always effective
   • DTo identify and exploit weaknesses in the AI defenses to improve their robustness
   Reveal Solution Discussion Select an answer first
6. 35application · medium

   A company uses an AI-based web application firewall (WAF) that blocks SQL injection attempts by analyzing query strings. A penetration tester wants to bypass the WAF to deliver a SQL injection payload. The tester has observed that the WAF model is trained on token sequences and is sensitive to unusual character combinations. Which technique is most likely to evade the WAF?

   • AEncode the payload using URL encoding and add benign tokens to break the model's token sequence.
   • BEncrypt the entire HTTP request using TLS to prevent the WAF from inspecting it.
   • CSend the payload in a single large packet to overwhelm the WAF's input size limit.
   • DUse a different HTTP method, such as POST instead of GET, to avoid detection.
   Reveal Solution Discussion Select an answer first