CT-GenAI ISTQB Certified Tester Testing with Generative AI (CT-GenAI) v1.0 Question and Answers

In the evaluation of GenAI outputs for testing, metrics must align with the specific nature of the task. Foranomaly identification, the goal is to correctly identify true issues without an overwhelming number of false positives; therefore,Precisionis the critical metric (the ratio of true anomalies to the total flagged). Conversely, forpartition testing(identifying valid and invalid input classes), the goal is thoroughness and ensuring no significant category is missed.Recallis the most appropriate metric here, as it measures the model's ability to "call back" or cover all possible relevant partitions from the requirement set. As highlighted in the CT-GenAI syllabus, evaluating AI effectiveness often requires a combination of these model-performance metrics. While "Accuracy" (Option D) provides a general view, it is often misleading in imbalanced testing scenarios (like anomaly detection where anomalies are rare). By using Precision and Recall together, a test organization can quantitatively assess if the AI is both trustworthy in its alerts and comprehensive in its test design coverage.

The most effective way to improve an LLM's performance on complex tasks likeimpact analysisis to provide a detailed, multi-stepInstructionorChain-of-Thoughtstructure. Option D is the best improvement because it breaks the "impact analysis" task into logical sub-tasks: mapping changes to modules, identifying related test cases, and prioritizing them based on risk and complexity. This structured approach guides the LLM through the "reasoning" steps a human expert would take, significantly reducing the likelihood of a superficial or incorrect analysis. While specifying a specialized role (Option B) or adding technical references (Option A) can help set the tone, they do not provide the model with the logical framework required to execute the task accurately. By explicitly defining theprocessthe LLM should follow, the tester ensures that the model evaluates the "depth" of the change rather than just listing files. This results in a more robust and actionable regression test suite, which is the primary goal of impact analysis in a modern software development lifecycle.

LLMOps(Large Language Model Operations) is the set of practices used to manage the lifecycle of LLMs in production. When an organization integrates an AI chatbot into its test processes, the primary operational concern ismaintaining data privacy and minimizing security risks, especially if using third-party APIs. Unlike traditional software, LLMs are "black boxes" that process every piece of data sent to them. A core LLMOps responsibility is ensuring that any "Prompt Data" (code, requirements, or logs) is not used by the provider to train their public models and that the communication channels are fully secured. While scalability (Option A) and latency (Option C) are important technical metrics, they are secondary to the catastrophic legal and reputational risk of a data breach. LLMOps in a testing context involves implementing data masking tools, monitoring for "Prompt Injection" attacks, and managing the "Grounding" data in vector databases to ensure it remains current and protected. This ensures the AI remains a safe and reliable asset within the enterprise testing ecosystem, rather than a liability for the organization’s intellectual property.

In professional Generative AI applications, thesystem prompt(sometimes called the system message) is the foundational set of instructions that defines the AI's persona, boundaries, and overall behavior. In a testing workflow, this is typically defined by atester or test engineerwho is configuring the AI assistant for a specific project. Unlike the user prompt, which changes with every interaction, the system prompt remains relatively static and acts as a "guardrail" to ensure the model stays in its role (e.g., "You are an expert in ISO 26262 automotive testing standards"). By defining the system prompt, the tester ensures that the model consistently uses specific terminology, adheres to data privacy constraints, and formats its output according to the team's requirements. While end users (Option B) provide the task-specific input, they do not usually have the permissions or technical need to alter the underlying system-level instructions. Similarly, while CI servers (Option C) might trigger the prompt, they do not "define" the human-centric logic contained within it. Properly crafting the system prompt is a core part of setting up an AI-augmented test environment.

This scenario describes a specialized form ofData Exfiltration(specifically targeting the model's internal "weights" or training memory). While data exfiltration usually refers to stealing data from a database, in the context of LLMs, it can also refer to techniques that force the model to "reveal" sensitive information it was trained on or data that exists within its current context window. By using long, repetitive, or specifically "crafted" prompts to overwhelm the model's normal attention mechanisms or safety filters, an attacker may cause the model to output verbatim snippets of proprietary information, PII, or internal documentation that should have remained confidential. This is different fromRequest Manipulation(Option D), which aims to change the model's behavior, orData Poisoning(Option A), which happens during training. In testing, this risk is high when models are fine-tuned on private company repositories. Testers must be aware that if a model is accessible to unauthorized users, those users might use adversarial prompting techniques to extract sensitive code or business logic through these types of data leakage attacks.

Generative AI is exceptionally capable of creating structured and unstructured data, but its role is limited to "generation" and "transformation," not infrastructure management or direct database administration. Creating production database backups (Option A) is a physical data management task involving the copying of actual stateful data from a server to storage; this is handled by database management systems (DBMS) and DevOps pipelines, not LLMs. Conversely, LLMs excel at the logic-based tasks listed in the other options. They can analyze requirements to identify and set boundary values (Option B) for input validation. They are also highly effective at creating combinatorial data (Option C), such as pairwise or all-combinations tables, by understanding the relationships between variables. Finally, one of the most powerful uses of GenAI in testing is generating synthetic datasets (Option D)—creating "fake" but realistically structured data that mimics production patterns without exposing Sensitive Personally Identifiable Information (SPII), thereby supporting privacy-compliant testing.

Few-shot promptingis the technique of providing a few examples (exemplars) within the prompt to demonstrate the desired task and output format to the LLM. In this scenario, providing 10 existing, high-quality test cases acts as a "pattern" for the model to follow. This is significantly more effective than "Zero-shot prompting" (Option D), where the model is given a task without examples and may deviate from the specific organizational format required (e.g., specific JSON structures or assertion styles). While "Prompt chaining" (Option A) is useful for breaking down complex tasks into sub-tasks, the primary need here is pattern recognition and replication, which is the core strength of Few-shot learning. "Meta prompting" (Option C) involves having the AI write the prompt itself, which is unnecessary when the team already has clear examples. By using Few-shot prompting, the tester "conditions" the model's latent space to prioritize the provided format, ensuring that all 500 generated test cases maintain consistency with the HTTP methods, headers, and assertion logic defined in the exemplars.

Hallucinations—where an LLM generates factually incorrect or nonsensical information—occur primarily when the model lacks sufficient specific information and "fills in the gaps" using probabilistic patterns from its training data. The most effective mitigation strategy is "grounding," which involves providing the model with detailed, project-specific context. By including technical specifications, existing API schemas, business rules, and identified constraints within the prompt, the tester restricts the model’s operational space to the "project realities." This ensures the model does not have to guess or improvise details about the System Under Test (SUT). In contrast, randomizing prompts (Option B) or relying on generic examples (Option C) increases the likelihood of inconsistent and inaccurate outputs. Furthermore, using "longer" or higher temperature settings (Option D) actually encourages creativity and randomness, which is the opposite of the precision required for testing and significantly increases the risk of hallucinations. Therefore, rich contextual grounding is the technical foundation for reliable AI-assisted test analysis.

Understanding the hierarchy of LLM types is vital for selecting the right tool for specific testing tasks.Foundation LLMsare trained on massive datasets to predict the next token; they excel at broad, creative "ideation" (Statement i) but often struggle with following specific instructions or constraints (making Statement iv incorrect).Instruction-tuned LLMshave undergone additional training (Fine-tuning) to follow explicit commands and templates. They are highly effective at structured tasks like converting requirements into Gherkin feature files (Statement ii).Reasoning LLMs(or those utilizing specialized prompting like Chain-of-Thought) are designed to handle complex, multi-stage logic. This makes them the superior choice for diagnostic tasks like root-cause analysis, where the model must synthesize information across logs and requirements to find a defect's origin (Statement iii). Statement v is incorrect because while instruction-tuned models are capable, complex "stepwise reasoning" usually requires specific prompting techniques or the inherent logic of specialized reasoning models. Therefore, the combination of i, ii, and iii represents the correct alignment of model capability to testing functionality.