README

Yuchen Shao^1,2, Yuheng Huang³, Jiawei Shen², Lei Ma^3,4, Ting Su², Chengcheng Wan^1,2*
1 Shanghai Innovation Institute
² Software Engineering Institute, East China Normal University, Shanghai, China
³ Department of Computer Science, The University of Tokyo, Tokyo, Japan
⁴ University of Alberta, Edmonton, AB, Canada

This is the artifact for our ICSE2025 paper "Are LLMs Correctly Integrated into Software Systems?". It is publicly archived at Figshare. This repository is made available under the Public Domain Dedication and License v1.0 whose full text can be found at: http://opendatacommons.org/licenses/pddl/1.0/ - see the LICENSE file for details. Notably, a misintegration case studied in our paper has been officially assigned CVE-2025-45150. If you have any questions, please feel free to contact me via email([email protected]).

We are grateful for the contribution made by our anonymous collaborator. Additionally, Mingyu Weng, Yiwen Sun, and Wenjing Liu have developed the Command-line Interface (CLI) to further enhance its functionality. The defect.csv file has been updated accordingly. You can review the latest updates!

Hydrangea is a defect library for LLM-enabled software. Hydrangea has 4 main petals, each corresponding to one of the major components where defects often arise: LLM agent, vector database, software component, and system.

What is LLM-enabled software?

It is software that integrates LLMs (large language models) with RAG (retrieval-augmented generation) support to realize intelligence features.

It contains four components:

1. LLM agent that manages LLM interfaces, constructs prompts, and invokes the LLM
2. Vector database that supports RAG algorithm and enhances the LLM agent
3. Software component that interacts with the first two components to perform certain tasks
4. System that manages resources and privileges to carry out the execution

What's inside the artifact:

For enhanced availability and reusability, we offer an organized defect library utilized in our manual studies.

Below are details of what is included in each part:

Application benchmark

A suite of 100 non-trivial projects that tightly integrates LLMs and vector databases in their workflow.

We have uploaded application.csv, which contains:

1. software project name
2. GitHub link and commit ID
3. classification
4. used LLM and vector database

Hydrangea Defect Library

The result of TABLE Ⅱ in our paper can be reproduced by this organized defect library.

In the uploaded defect.csv, we have documented different cases for the same defect type, as defects can manifest in various ways. For each distinct case of the same defect, we have separated them with a blank line and labeled them as case 1, case 2, and so on, according to the specific circumstances.

It contains:

A collection of defects in these projects (involves 100 projects), containing

1. the defect type and its detailed explanation
2. the exact file and source-code line location of the defect
3. the consequences of defect
4. the defect-triggering tests

The meaning of different columns in defect.csv:

1. APP: the applications from GitHub.
2. commit url: the relevant version of the application on GitHub.
3. types: different defect types.
4. cases: different examples for each defect type. Cells containing a "/" indicate that there is only one case for that defect type.
5. explanation: details of the defect.
6. consequences: the impacts of the defect. Here we use the abbreviations: ST refer to fail-stops, IC refer to incorrectness, SL refer to slower execution, UI refer to unfriendly user interface, TK refer to more tokens, and IS refer to insecure.
7. source-code locations: The location of the code file where the defect occurs.
8. defect-triggering tests: The software input that triggers the defect.

Quick Start-How to use Hydranger?

Take LocalAGI as an example. It makes plans to guide users to achieve their goals. However, due to its infinite loop design with time intervals, it repeatedly refines a subset of the generated steps, without providing a final version that contains all the refinements. Making things worse, this loop could only be broken by terminating the entire application, significantly degrading user experience.

Tutorial

1. Open application.csv to find the corresponding GitHub link and commit ID for this application.
2. Review defect.csv to get an overview of the defect and the associated defect-triggering tests.
3. You can attempt to reproduce the issue.

Command-line interface: Hydrangea command

🚀 Setup

1. Clone Defects4J:

   git clone https://github.com/ecnusse/Hydrangea.git

2. install dependencies

   pip install -e .

🎯 Command Overview

Command	Description	Main Parameters
apps	📱 List all applications, supports multi-dimensional filtering	--classification, --llm, --vdb
bids	🐛 List all defect IDs, supports filtering by application	--app
info	📊 Display metadata information of a specific defect	app, bid
test	🧪 Display test information	app, bid, --trigger

---

1. 📱 apps Command — List Applications

Basic Usage

# List all applications
hydrangea apps

# View detailed help information
hydrangea apps --help

# Filtering application based on llm
hydrangea apps --llm OpenAI

# Filtering application based on vector database
hydrangea apps --vdb chroma

# Filtering application based on language
hydrangea apps --language python

---

2. 🐛 bids Command — List Defect IDs

Basic Usage

# List all defect IDs
hydrangea bids

🔍 Filter Defect IDs by Application

# List all defect IDs for a specific application
hydrangea bids --app LocalAGI

💡 Tip: Application names support fuzzy matching. All applications containing the specified keyword will be listed with their defect IDs.

---

3. 📊 info Command — View Detailed Defect Information

Basic Usage

# View detailed information for a specific defect
hydrangea info "camel" camel-ai-camel-privacy_violation-/
hydrangea info MEMGPT cpacker-MemGPT-exceeding_llm_content_limit-case1

📋 Output Information Includes

• Defect Type
• Application Basic Information
• Defect Description and Impact

---

4. 🧪 test Command — Display Test Information

📝 Parameter Description

Parameter	Type	Description
app	Required	Application name
bid	Required	Defect ID
--trigger	Optional	Show detailed trigger test information

🚀 Usage Examples

# Show basic test information
hydrangea test camel camel-ai-camel-privacy_violation-/

# Show detailed trigger tests
hydrangea test camel camel-ai-camel-privacy_violation-/ --trigger

📤 Output Examples

🔸 Basic Mode Output

Test information for camel - camel-ai-camel-privacy_violation-/
Defect type: privacy violation
Case: /
Use --trigger to see detailed trigger tests

🔸 Trigger Test Mode Output

trigger_tests:
- 1.Set up Camel: Ensure the project is correctly set up in your local environment.
2.Configure Roles: Set up a role-playing scenario where one agent has restricted privileges (e.g., "Guest User") and another has administrative privileges.
3.Define Security Boundaries: Establish clear system message constraints about access permissions.
4.Execute Test: Initiate conversation where the restricted role attempts to access privileged operations.
5.Observe Behavior: Monitor if the AI system properly enforces access controls or inadvertently reveals sensitive information.

---

⭐ If this project is helpful to you, please give us a Star!

Made with ❤️ by Ungifted77,Evensunnn,SunsetB612