Data for: Abusability of automation apps in intimate partner violence
Data files
Jul 29, 2025 version files 108.41 MB
-
final_labels_pseudonymized_csv.zip
50.87 KB
-
human_label_exceed.zip
15.77 KB
-
model_labeled_rationale.zip
107.27 MB
-
README.md
3.55 KB
-
script_for_functionality.zip
1.07 MB
Abstract
The research investigates how popular automation apps—Apple Shortcuts, Tasker, IFTTT, and Samsung Modes & Routines—can be exploited maliciously in intimate partner violence (IPV) contexts. The study systematically examines the capabilities of these apps to facilitate surveillance, control, and harassment. We designed an LLM-based detector to investigate whether an iOS Shortcut recipe is capable of performing a surveillance, lockout/control, overloading, or impersonation attack. The detector takes the iCloud URL of shared Shortcuts, parses that into an LLM-readable format, and analyzes using LLM-detector. In total, we evaluated 12,962 publicly available Apple Shortcut recipes to identify those capable of IPV abuse.
Dataset DOI: 10.5061/dryad.b2rbnzssm
Description of the data and file structure
This dataset includes our LLM detector, data preparation scripts, and evaluated Shortcut recipes about IPV (Intimate Partner Violence) attacks.
Files and variables
File: final_labels_pseudonymized_csv.zip (CSV files)
Description: includes all CSV files of trinary results of whether a recipe can cause IPV harms, labeled by LLM. If you need the mapping from pseudo-names to real names, please contact us. All files include the following three fields:
- Input File: string, the pseudo-mapped names of online Shortcut recipes.
- Action: one from overload, impersonation, spy, and lockout. Indicate which attack operation is evaluated by the LLM for that recipe.
- Response: one from yes, no, or maybe. Yes means the LLM considers this recipe contains 'Action' attack, no is vice versa. Maybe means the LLM considers this recipe to have the potential of conducting an IPV attack.
File: human_label_exceed.zip
Description: includes the human labels of shortcut recipes that are too long for LLM to evaluate. This zip file contains three folders of CSV files, each folder denotes one domain we evaluated. There are only two columns in each file:
- Recipe: string, the name of online Shortcut recipes that are too long for LLMs to evaluate
- Result: Yes/No, labelled by human, identify if this Shortcut recipe could conduct an IPV attack.
File: model_labeled_rationale.zip
Description: includes all models evaluated rationales, used to make a decision on whether a shortcut recipe is capable of conducting an IPV attack. This zip file contains five folders, each include the 4 attack types & one re-run of impersonation using GPT4o.
Each file in these folders is a json file, include the full history of conversation with LLMs: the history includes the beginning system prompts, the recipe contents, and the LLM responses. The prompts for each attack can be find in script* * for functionality.
File: script_for_functionality.zip
Description: includes how to turn an iCloud link/iCloud links of iOS Shortcut recipes into XML files (credit to Glenn 'devalias' Grant, https://gist.github.com/0xdevalias/27d9aea9529be7b6ce59055332a94477); YAML files of Shortcut actions generated from XML files; code filter to check if a YAML file needs to evaluated by LLM detector; token calculator to check if the YAML file can be evaluated by LLM detector; and LLM detector to determine if a recipe are capable of conducting IPV attacks. Please refer to the README.md after unzipping this file.
For the two zip file below, please contact the author directly.
File: marketed_YAML.zip
Description: Includes the YAML files of shortcut recipes that are marketed for abusive keywords. This zip file contains 4 folders, each is one of the IPV attacks, in each folder, each file is a txt file that contains the text representation of Shortcut recipes that are martketed for harming.
File: evaluated_yaml_files.zip
Description: Includes the YAML files of shortcut recipes we get from four domains. This contains the text representation of recipes from four domains, these data is generated from the raw XML files through functionality provided by script_for_functionality.zip. For safety reason we do not provide the raw XML files.