Watermarking has become a crucial technique for ensuring provenance and accountability in AI-generated source code. As large language models (LLMs) are increasingly integrated into development workflows, reliable attribution remains challenging. In practice, most developers rely on commercial LLM APIs operating under black-box constraints, making existing approaches that require access to the decoding process less feasible for real-world integration. To address this limitation, we propose PromptMark, a black-box, prompt-guided watermarking framework that embeds invisible yet statistically detectable signals into generated code via structured input instructions. The method steers models toward subtle identifier and comment naming patterns while preserving the functional correctness and structural integrity of the generated code. Detection is performed using statistical tests designed to remain reliable across varying code lengths and model outputs. The embedding is further refined through an iterative feedback loop, where prompts are updated based on watermark detection scores. Experiments on the MBPP and HumanEval benchmarks show that PromptMark consistently achieves strong watermark detectability while maintaining high code correctness, outperforming baseline approaches.

Black-box Test Suite Minimization (TSM) aims to remove redundant test cases from a test suite while retaining its fault detection capability without analyzing the production code. This makes it particularly efficient to be used in industry projects. These techniques utilize metrics such as test history, commit complexity or test code diversity (or similarity) to guide test case selection. Similarly, change-proneness (CP) can also guide to successful test case selection as it indicates the likelihood of having faults. We propose CP as a metric for TSM and implement into an approach, Change-proneness based Test suite Minimization (CTM), where relationships between test cases and their depending classes are measured using CP values. CTM first calculates class-level CP, followed by identifying the dependency between the test cases and classes. Then the association between test cases and their related classes are calculated using various statistical measures such as geometric mean etc. Finally test cases with the highest association values are selected. To demonstrate the effectiveness and efficiency of CP as a TSM metric, we compared CTM with state of the art, AST-based Test case Minimizer (ATM) on 15 Java projects with 617 versions. The experimental results demonstrate that CTM achieves the same average fault detection accuracy as ATM (0.67) while reducing execution time by 114.5 folds.

Dark patterns manipulate user choices through deceptive UI tactics. Any automated detection technique for dark patterns must address the varying nature of dark patterns. Existing detection techniques need manual intervention in some cases. In other cases, techniques are not generalized due to overfitting problems; for example, these techniques can not handle cases where texts are semantically similar but possess lexical differences. We propose an automated dark pattern text detection technique that is generalized. We synthesize inclusive definitions of dark pattern categories. This contextual information is prioritized using in-context learning capabilities of GPT-3 to detect and classify dark pattern texts. Results show that our technique offers satisfactory performance for 6 out of 7 dark pattern categories explored in this study. We also validate the improved generalization capability of our technique by outperforming an existing baseline model on a test dataset.

Automated End-to-End (E2E) web testing is a key component in modern rapid development to validate system functionality. However, there are no resources supporting practitioners on how diverse scenarios are tested manually. This paper presents WebEV, a dataset containing E2E test cases from open-source popular projects. Projects are selected based on - i) Cypress-based automation, ii) popularity on GitHub and iii) executability of test cases. The dataset contains information regarding each test command along with the incurred state change representation. Snapshots of the application are used to retrieve - i) the current URL of the application, ii) the screenshot and HTML text of the entire page, and iii) the screenshot and HTML text of an operated UI element. This process is done both before and after each command execution to capture the perception of testers on each state transition, i.e., extract their thought process during testing. This dataset can assist the research community to model user web interaction, predicting the tester’s perception, and improving the state of automated testing approaches. Moreover, WebEV can be used to mine how automated approaches differ from real-life E2E test scenarios.

Web application failure detection relies mostly on the tester’s creativity, leaving test automation to only ease executing repetitive tasks. Existing automated testing techniques opt for test path diversity or input generation but not the tester’s behavioral patterns. For example, testing deeply nested business logic, proper form submission, or non-redundant navigation are not considered. This paper proposes eBAT, an automated testing approach that considers those testers’ interaction patterns from observation. A behavior-driven action selection strategy is derived from these patterns to interact with the system. Actionable elements (buttons, links, inputs, etc.) obtained through state abstraction and interaction pattern-wise grouping are operated in a tree-based manner. The effectiveness and efficiency of eBAT are evaluated as the unique number of failures detected and the detection rate respectively. Results compared against the state-of the-art indicate significant improvement in failure detection with similar code coverage. Moreover, eBAT outperforms the baseline failure detection rate in 5 out of 6 benchmark projects.