hujk/OGAAA

Files

HenryChou020514 01bb07dba8 Flatten 1_raw_dataset submodules into plain tracked files

FocalLoRA, Should-It-Be-Executed-Or-Processed, and topicattack were
nested git repos (with an inner FocalLoRA/data/FocalLoRA/.git as well).
Drop their .git history and track the contents directly in this repo
instead of as submodules/gitlinks.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

2026-07-07 19:06:09 +08:00

data

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

injectagent

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

prompts

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

attack_defense_tools.py

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

chatbot.py

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

construct_attack.py

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

README.md

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

run_baselines.py

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

shield.py

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

utils.py

Flatten 1_raw_dataset submodules into plain tracked files

2026-07-07 19:06:09 +08:00

README.md

TopicAttack: An Indirect Prompt Injection Attack via Topic Transition

Official code implementation for EMNLP 2025 accepted paper: TopicAttack: An Indirect Prompt Injection Attack via Topic Transition (https://arxiv.org/pdf/2507.13686)

Environment

conda creat -n topicattack python=3.10
conda activate topicattack
pip install openai

Chatbot Evaluation

We have crafted the dataset with TopicAttack and you can find them in data/crafted_instruction_data_squad_conversation_attack_complete.json and data/crafted_instruction_data_tri_conversation_attack_complete.json. To evaluate the performance, you can run the following code :

python run_baselines.py \
    --victim_model_path Qwen/Qwen2-7B-Instruct \
    --data_path data/crafted_instruction_data_squad_conversation_attack_complete.json \
    --victim_system_path prompts/generator_system_prompt.txt \
    --attacks conv_attack \
    --defenses none sandwich  \
    --log_file squad_logs/sandwich-conv-ignore-qwen2-7b-squad-complete.txt

You can also replace the victim model with other models such as GPT-4o.

Attack Construction

To construct the attack data, please put your OpenAI api key into the chatbot.py. Then you can run the following code:

python construct_attack.py \
    --model gpt-4o \
    --data_path ./data/crafted_instruction_data_squad_injection_qa.json \
    --template_path prompts/attack_prompt_transfer.txt \
    --chatbot_system_path prompts/chatbot_system.txt \
    --output_path  ./data/crafted_instruction_data_squad_conversation_attack_complete.json

Agent Evaluation

Please step into the injectagent directory for more details, we implement our attack based on InjecAgent.